Jan K. Sykulski, Richard L. Stoll (Auth.), Jan K. Sykulski (Eds.) - Computational Magnetics-Springer Netherlands (1995)

Computational Magnetics
Computational Magnetics
Editor
Jan K. Sykulski
E3
SPRINGER-SCIENCE+BUSINESS MEDIA, B. V.
First edition 1995
© 1995 Springer Science+Business Media Dordrecht

Originally published by Chapman & Hall in 1995
Softcover reprint of the hardcover 1st edition 1995
Typeset in 10/12pt Times Roman by Colset Pte Ltd, Singapore
ISBN 978-94-010-4557-5 ISBN 978-94-011-1278-9 (eBook)

DOI 10.1007/978-94-011-1278-9
Apart from any fair dealing for the purposes of research or private study, or
criticism or review, as permitted under the UK Copyright Designs and Patents
Act, 1988, this publication may not be reproduced, stored, or transmitted, in
any form or by any means, without the prior permission in writing of the
publishers, or in the case of reprographic reproduction only in accordance with
the terms of the licences issued by the Copyright Licensing Agency in the UK,
or in accordance with the terms of licences issued by the appropriate
Reproduction Rights Organization outside the UK. Enquiries concerning
reproduction outside the terms stated here should be sent to the publishers at the
London address printed on this page.
The publisher makes no representation, express or implied, with regard to the
accuracy of the information contained in this book and cannot accept any legal
responsibility or liability for any errors or omissions that may be made.
A catalogue record for this book is available from the British Library
(^) Printed on permanent acid-free text paper, manufactured in
accordance with ANSI/NISO Z39.48-1992 and ANSI/NISO Z39.48-1984

(Permanence of Paper).
Contents
Preface ix
1. A Brief Review of Magnetic Fields

Jan K. Syku/ski and Richard L. Stoll
1.1 Maxwell's equations 1
1.2 Scalar and vector potentials 4
1.3 Classification of fields 8
1.4 Boundary conditions 9
1.5 Magnetic materials 12
1.6 Energy storage and forces in magnetic field systems 15
1.7 Eddy-current skin effect 18
1.8 Geometrical structure of electromagnetic fields and
differential forms 21
1.9 The method of tubes and slices 24
1.10 Field calculations using tubes and slices 29
1.11 Preparatory and further reading 32
2. Finite-Difference Methods 35
Richard L. Stoll and Kazimierz Zakrzewski
2.1 Introduction 35
2.2 Discretization, finite-difference operators and truncation
errors 36
2.3 A one-dimensional parabolic finite-difference equation 37
2.4 Convergence, compatibility and stability 42
2.5 Convergence of nonlinear diffusion equations 44
2.6 Finite-difference methods applied to magnetic field
problems 48
2.6.1 Steady-state eddy currents in ferromagnetic plates,
laminations or cylinders 49
2.6.2 Treatment of magnetic nonlinearity and hysteresis 52
2.6.3 Transient one-dimensional solutions 55
2.6.4 Two-dimensional static solutions in multi-region
problems 57
2.6.5 More two-dimensional static and steady-state
solutions 63
vi Contents
2.6.6 Two-dimensional transient or nonlinear

steady-state problems 70
2.6.7 Three-dimensional solution of Laplacian and
Poissonian problems 73
3. Finite-Element and Boundary-Element Methods 78

Ryszard Sikora
3.1 Introduction 78
3.2 Weighted residuals 78
3.3 The finite-element approach 82
3.4 Isoparametric elements 90
3.5 Nonlinearity in statics 93
3.6 Linear time-dependent problems 100
3.7 Nonlinear time-dependent problems 104
3.8 Boundary-element method 109
3.9 Open-boundary problems 113
3.10 Hierarchical elements 123
3.11 Three-dimensional problems 125
3.12 Numerical solution of large systems of equations 130
3.13 Applications 135
3.13.1 Electromagnet 135
3.13.2 Leakage field in a transformer 137
4. Reluctance Networks 145

Janusz Turowski
4.1 A survey of existing solutions and programs 145
4.2 Reluctance networks and finite differences 148
4.3 Leakage fields in transformers 150
4.4 Computational aspects 161
4.5 Interactive design and detection of 'hot spots' 164
4.6 Testing RNM on other electromagnetic devices 171
4.7 Reluctance networks and finite elements 171
4.8 Time-dependent fields 174
4.9 Experimental verification of the method 176
5. Field Synthesis 179

Krystyn Pawluk
5.1 Introduction 179
5.1.1 What is field synthesis? 179
5.1.2 Development of methods and techniques 181
5.2 A survey of integral equations 182
5.2.1 Types of equations 182
5.2.2 Green's functions 186
5.3 Types of problems 196
Contents vii
5.3.1 Inverse problem to Laplace's equation 196

5.3.2 Inverse problem to Poisson's equation 199
5.3.3 Variable boundary problem 201
5.3.4 Synthesis of material parameters 203
5.4 Solution of synthesis problems 204
5.4.1 General remarks 204
5.4.2 Quadrature 204
5.4.3 Direct discretization 208
5.4.4 Conditions for solutions 210
5.4.5 Least-squares methods 214
5.4.6 Regularization 219
5.4.7 Parameter of regularization 222
5.5 Examples of solved problems 224
5.5.1 Synthesis by searching for boundary conditions 224
5.5.2 An air-cored coil 227
5.5.3 Boundary conditions for a transmission line 229
5.5.4 Other problems 232
6, Coupled Fields 234

Janusz Turowski
6.1 Types of coupled fields. Strong and weak coupling 234
6.2 Electro-magneto-mechanical fields 235
6.2.1 Hamilton's principle 235
6.2.2 Euler's equation 238
6.2.3 Euler-Lagrange equation 240
6.2.4 Static electric and magnetic fields in isotropic
media 241
6.2.5 Maxwell's equations as a consequence of
Hamilton's principle 243
6.2.6 Generalized vector of power density of coupled
fields 244
6.2.7 Field of forces 245
6.2.8 Vibration of busbars and conductors 250
6.2.9 Maxwell's stress tensor 252
6.3 MHD 254
6.4 Non-homogenous media 260
6.5 Magnetostriction 261
6.6 Coupled magnetic and thermal fields 261
6.6.1 Magnetoelasticity 261
6.6.2 Magnetothermoelasticity 265
6.6.3 Thermokinetics 267
6.6.4 Thermometric method of power loss measurement 268
6.6.5 Local overheating criteria 272
6.6.6 Induction heating 276
6.7 Levitation 281
viii Contents
7. Computer Aided Design in Magnetics 285

Jan K. Sykulski
7.1 The CAD environment 285
7.2 Pre-processing and mesh generation 289
7.3 Code development and limitations 297
7.4 The art of sparse matrices 298
7.5 The ICCG method 303
7.6 Error analysis and adaptive meshing 305
7.7 Optimization 313
7.8 Post-processing and field visualization 316
7.9 The impact of parallel computing 321
7.10 A short survey of CAD systems for magnetics 323
8. Experimental Methods 330

Kazimierz Zakrzewski
8.1 Introduction 330
8.2 Physical modelling 330
8.2.1 Scaling criteria 330
8.2.2 Modelling fields and losses in ferromagnetics 332
8.2.3 Linear motors 335
8.3 Experimental techniques 338
8.3.1 Experimental models 338
8.3.2 Basic measurements 339
8.3.3 Flux and power loss in magnetic cores 341
8.3.4 Measurements in air-gaps of electrical machines 342
8.3.5 Power loss in transformer tanks 343
8.3.6 Leakage fields 344
8.4 Experiments for coupled fields 345
8.4.1 Forces and stresses 345
8.4.2 Temperature fields 347
8.4.3 Thermometric method 347
8.5 Analogue modelling 348
8.6 Experimental errors 352
References 354
Index 371
Preface
Computational Magnetics comprises a spectrum of subjects covering field and

potential theory, interpolation and approximation techniques, numerical
methods and programming routines, as well as relevant experimental
procedures and analogue systems. Its objective is to provide a working
environment and a set of tools for a designer of magnetic apparatus, or a
system analyst, to facilitate development of appropriate models of devices and
processes for the purpose of simulation and to establish equivalent circuits,
or other simplified representations, and their parameters. It is a practical tool
that aims at helping an engineer in the difficult task of designing magnetic
devices and optimizing their performance. Moreover, it aids an understanding
of the behaviour of such devices.
Advances in computational magnetics stem from progress in various areas
of mathematics, physics and computer science, and, in turn, feed back and
contribute to _developments in field and potential theory, calculus of
variations, optimization methods, numerical analysis, and computational
geometry. Scientists, engineers and software developers allover the world
actively pursue research in many diverse topics relevant to computational
magnetics, ranging from gauge theory, scalar and vector potential
formulations, and methods of synthesis to benchmarking techniques, parallel
processing, computer graphics and data structures. Error estimation and
adaptive solver technology, treatment of exterior circuits and coupling of
magnetic to other field systems are also very important.
The notion of computer-aided design (CAD) is central to contemporary
systems for magnetic analysis and provides the framework for implementation
of key concepts of computational magnetics. Numerous commercial CAD
systems are now available for magnetics work and the software has grown to
playa fundamental role in the industrial design environment. As a result the
designer's approach to problems is inevitably changing too, as conceptual
models used in design are now influenced by what is offered by these new
powerful tools. New questions arise however on inherent simplification of
models and methods used, relevance of computed quantities to the needs of
the designer and accuracy of computation. Another important factor is the
convenience with which the solution is found, that is the user-friendliness of
the program, and how the software communicates with other components of
the design system, Le. versatility and compatibility of data structures. And
x Preface
last but not least the user of any CAD software will be looking very carefully
at the total time required to set up the problem, analyze it and interrogate the
solution. Modern computers are immensely powerful and fast, but complexity
of some real engineering problems, especially if large three-dimensional
structures or optimization are involved, may be such that resulting computing
times are impractical for design purposes. Approximations and simplifications
are then sought so that a compromise may be reached between accuracy and
speed of computation. Inter-active design becomes particularly helpful where
the designer is in constant 'conversation' with an appropriate software system
and thus his knowledge and expertise are directly and immediately used when
required.
There is an almost bewildering range of computational methods in mag-
netics, each with a vast literature, well developed theory and dedicated
supporters. The reader is no doubt familiar with some of the methods, such
as separation of variables, Laplace transforms, conformal transformations,
analogue methods (e.g. transmission-line modelling) or images. Numerical
methods have been found particularly helpful for CAD applications and
they have indeed dominated the scene in recent years. The most important
techniques, all discussed in this book in some depth, are finite differences,
finite elements, boundary elements, reluctance networks (based on circuit
analogy) and tubes and slices (a graphical computational scheme based on
dual energy approach). Although all these methods aim at describing the same
physical processes, as defined by appropriate field equations, they differ in
both the formulation and practical implementation. A typical CAD system
will be firmly rooted in one of such methods and will have an appropriate
graphical user interface and possibly a design or optimization shell, as well
"as some other specialist toolboxes.
Methods of field synthesis are being introduced gradually into computa-
tional magnetics, as many of practical engineering designs aim at achieving
specific characteristics or particular field levels or distributions. Although
ultimately a finite-element technique or other method of analysis is usually
employed, the mathematics of field synthesis is somewhat different and
requires careful examination. Solutions of such inverse problems are increas-
ingly of interest to designers.
This book has been written by a group of authors who are all actively
engaged in research and pursuing developments of various aspects of com-
putational magnetics. The coverage is fairly comprehensive and addresses a
number of fundamental issues, some of which have been mentioned above.
We do not concentrate on any particular method, nor do we dwell on theory
alone. The flavour of the book is the link between underlying mathematical
and physical concepts and computational practice, with particular emphasis
on engineering applications. We have thus tried to bridge the gap between
theory, available computational techniques and engineering and design
requirements. It is up to the reader to decide whether we have achieved our
objective.
The reader is assumed to be familiar with fundamental concepts of electro-
Preface xi
magnetics as covered by a typical second year undergraduate course. A

recently published textbook and an accompanying computer disk may be
helpful in revising the subject (P. Hammond and J.K. Sykulski, Engineering
electromagnetism: physical processes and computation, Oxford University
Press, New York, 1994, ISBN 0 198562896 or ISBN 0 19856288 8). Con-
versely, a reader who wishes to further his studies of a particular method or
technique, beyond the coverage offered in one of the chapters, or to learn
about other methods not discussed in this book, will find a very extensive list
of references of both specialist books and monographs, as well as scientific
papers published in learned societies journals and conference proceedings.
There are some conferences (e.g. COMPUMAG, ISEF) which regularly report
on advances in computational magnetics and are therefore particularly useful
to those who are interested in the subject.
Our book presents the state of the art in computational magnetics.. The
subject is vast and it would be impossible to treat equally thoroughly every
relevant topic. We have thus selected and presented in greater depth some
of the issues, but only briefly mentioned or referred the reader to relevant
literature in case of others. We think that the book will be particularly useful
to practising engineers and designers in various sections of industry, in parti-
cular electromechanical, dealing with magnetic devices at low frequency, large
and small, and to researchers and postgraduate students involved in magnetic
analysis. It will also be helpful in running postgraduate and continuing
education courses and to final year undergraduate students taking specialist
options and for project work.
The book starts in the first chapter with a brief review of magnetic fields,
covering Maxwell's equations, scalar and vector potential formulations,
classification of fields, boundary conditions and magnetics materials. It then
goes into the discussion of energy storage and forces and gives a brief account
of the eddy-current skin effect. Finally, it introduces the geometrical structure
of fields and the notion of differential forms, which give rise to a geometrical
computational technique known as a method of tubes and slices.
The next three chapters discuss theory, computational aspects and applica-
tions of the most popular numerical methods. Chapter 2 deals with the oldest
and at one time hugely successful method of finite differences. In its simplest
form the method is easily understood and almost intuitive; it is still very
useful for special purposes and when the geometry is simple. The discussion
concentrates firstly on general matters of stability and convergence of the
numerical scheme, and then on more specific questions of nonlinearity,
hysteresis, multi-region problems and transient solutions. Many practical
solutions are demonstrated. Chapter 3 describes the two techniques which
have completely dominated the scene and are used in almost all commercial
software: the finite-element and boundary-element methods. The algorithms
are much more complicated than the finite-difference formulation. A com-
prehensive derivation of basic theory is followed by some of the more
advanced features, such as nonlinearity, open-boundaries and hierarchical
elements. Some examples of practical solutions are given. Finally, Chapter 4
xii Preface
presents a rather unique implementation of an old approach based on equi-

valent reluctance networks. With the help of a dedicated computer program
the method offers great economy of computation, providing approximate
answers in a matter of seconds, even for complicated three-dimensional
structures. The method is partly empirical and requires experimental verifi-
cation, but numerous successful solutions have been obtained and are
reported, together with a more general discussion of theory and applications.
Chapter 5 deals with field synthesis. After defining the term it then provides
a survey of integral equations and discusses types of typical synthesis tasks:
inverse problems to Laplace's or Poisson's equations, variable boundary
problems and synthesis of environmental parameters. Possible methods of
solution are reviewed and discussed in some depth, including one of the most
promising techniques of regularization. Some practical examples are also
described. Coupling between magnetic and thermal, mechanical and other
field systems is the subject of Chapter 6. This is an area of growing interest
to engineers as most design problems involve interaction between different
types of fields. Relevant theory is explained and various mathematical models
are put forward. The approach is very practical and numerous examples are
included. Consequences of strong and weak coupling are carefully studied and
possibilities of approximations and simplifications of models and equations
are explored.
Chapter 7 addresses the fundamental issues of computer aided design in
magnetics and attempts to summarize the state of the art in both theory
and implementations. The CAD environment is introduced together with
its main components: preprocessor, solver and post-processor. These are
discussed in some detail and many recent developments are highlighted. The
discussion focuses on such topics as error analysis and adaptive meshing,
sparse matrices and compact storage schemes, optimization, field visualization
and parallel computing. A short survey of currently available commercial
CAD systems for magnetics is provided for information. Finally, Chapter 8
looks at experimental methods. Fundamentals of physical modelling and
scaling criteria are explained, followed by a discussion of various experi-
mental techniques, and concluded with classification of experimental errors.
Measurement is a very important component of a modelling process, both as
a verification of theoretically derived results and as an investigating tool in
its own right.
As the editor and co-author of this book I have been privileged to cooperate
with my distinguished colleagues and I would like to express my gratitude for
their innovative approach to the subject and all the hard work they put in to
complete the manuscript and meet the deadlines. The book is a collaborative
effort of the following group of authors:
• Krystyn Pawluk - Professor of Electrical Engineering, Institute of Electro-

technics, Warsaw, Poland
• Ryszard Sikora - Professor and Head of Department of Electrical Engi-
neering, Technical University of Szczecin, Poland
Preface xiii
• Richard L. Stoll - Reader in Electrical Machines, Department of Electrical

Engineering, University of Southampton, UK
• Jan K. Sykulski - Lecturer in Electrical Machines, Electromagnetism and
CAD at the Department of Electrical Engineering, University of Southamp-
ton, UK, and Visiting Professor at the Technical University of Lodz, Poland
• Janusz Turowski - Professor and Vice-Rektor of the Technical University
of Lodz, Poland
• Kazimierz Zakrzewski - Professor and Head of the Institute of Electrical
Machines and Transformers, Technical University Of Lodz, Poland
The collaboration between the Department of Electrical Engineering at

Southampton, UK, and the Universities of Lodz, Szczecin and Institute of
Electrotechnics in Warsaw, Poland, has a long history and the research links
have been active for over fifteen years. The International Symposium on
Electromagnetic Fields in Electrical Engineering (ISEF), a biannual event
held alternately in Poland and another European venue, was hosted by the
University of Southampton in 1991 and was a very successful meeting. The
ISEF Symposium is one of the venues where advances in computational
magnetics are regularly reported. The present book reflects the strength of
the cooperation between researchers and institutions in the two countries.
I hope that the reader will find the book enjoyable and helpful. Suggestions
for future changes and improvements will be most welcome.
Jan K. Sykulski
Southampton
May 1994
1
A Brief Review of Magnetic Fields
Jan K. Sykulski and Richard L. Stoll
1.1 MAXWELL'S EQUATIONS

Most of this book addresses the problems of how to find, estimate or compute
the distribution of magnetic fields for different geometries and under v.arious
particular conditions. Such fields are described, in a most compact form, by
Maxwell's equations, and thus for most of the time we shall be seeking
solutions to these equations. Readers familiar with the theory may wish to
proceed directly to a chapter dealing with a particular method or technique
of interest, but we shall nevertheless start by introducing briefly the most
fundamental aspects of the underlying theory and basic formulations.
Maxwell was the first investigator to use partial differential equations to
describe electromagnetic phenomena. For general time-varying fields, using
today's notation, Maxwell's equations may be written in differential form as
aD
curlH = J + at (1.1)
curlE
oB
= -- (1.2)
at
divD = p (1.3)
divB =0 (1.4)
where:
H magnetic field intensity (Am-I);
E electric flux intensity (V m -I);
B magnetic flux density (Wb m- 2, or T);
D electric flux density (C m- 2);
J electric current density (A m -2);
p electric charge density (C m -3).
The current density J and the charge density p are the sources of the field and
are related through the equation of continuity, which can be written as
divJ = _ iJp (1.5)

ot
and thus specifies the conservation of charge.
2 A Brief Review of Magnetic Fields
The above equations are supplemented by the constitutive equations

describing macroscopic properties of the medium
D =EE (1.6)
B=~H (1.7)
J = aE (1.8)
where the constitutive parameters E, ~ and a denote, respectively, the per-
mittivity (F m -I), permeability (H m -I), and conductivity (S m -I) of the
medium. For isotropic media these parameters are scalars, but for anisotropic
materials they become tensors. Moreover, for non-homogeneous materials,
they are functions of position.
For static fields, where the field quantities do not vary with time, eqns (I. I ),
(1.2) and (1.5) reduce to
curlH = J (1.9)
curiE=0 (1.10)
divJ = 0 (1.11)
which shows that there is no interaction between electric and magnetic
fields, and thus we can have separately an electrostatic case or a magnetostatic
case.
An important special condition arises also when all field quantities vary
sinusoidally with time as e jOlt , where 00 is an angular frequency. Using phasor
notation, the Maxwell's equations involving time derivatives can now be
written as
curlH = J + jooD (1.12)
curiE = - joo.D (1.13)
divJ = -joop (1.14)
The three vector operators grad, div and curl are very helpful in all of
the above compact notations. The definitions of these operators are as
follows:
a<l>
grad<l>
A
=-0 (1.15)
an
and is thus the total or maximum slope of the scalar 41, where the unit
vector points in the 'uphill' direction normal to the equipotential surface.
Secondly,
div F = lim ~ t F . ds
v-o v 1
(1.16)
describes the net outflow of vector F per unit volume, for a small volume,
and finally
Maxwell's Equations 3
curl F = lim ~ " F . dl (1.17)

S~O s r
is the circulation per unit area, for a small area. In a rectangular coordinate
system (x, Y, z) it is convenient to introduce a nabla (sometimes called a del)
operator, defined as
(1.18)
and the three operations are reduced to

= V
grad (1.19)
divF = V· F (1.20)
curlF = V xF (1.21)
which also give alternative notation. In other coordinate systems, definitions
of eqns (1.15), (1.16) and (1.17) have to be obtained directly, but the nabla
notation is still often used, although it will no longer imply operation of
eqn (1.18).
Another form of the equations may be found by application of some
fundamental integral relationships of vectors. The two most important are
Gauss's or divergence theorem
JJJ V . F dv = ; F . ds (1.22)
and Stoke's or circulation theorem
HVXF.ds=;F.dl (1.23)
Hence, for example, for a steady current flow the integral equation equivalent
to eqn(1.11) will read
'J.
s
ds =0 (1.24)
Equations (1.11) and (1.24) contain the same information and are vector
equivalents of the first Kirchhoffs current law.
In electrostatics we can write
'O.ds=Q (1.25)
s
Thus, Gauss's theorem states that the electric flux over any arbitrary closed
surface is always equal to the strength of the enclosed sources. The last
equation is equivalent to the statement of eqn (1.3).
In magnetostatics there may be no net polarity as all magnets consist of
dipoles, so that
;B'ds=O (1.26)
s
which in differential form is expressed by eqn (1.4). Moreover, application of
Stoke's theorem to eqn (1.9) yields
;H'dl=HJ'ds=I (1.27)
s
which is known as Ampere's equation and shows that the magnetic field of
a current is non-conservative.
Faraday's law, eqn (1.2), can also be written as
.i E. dl = _ d~ (1.28)
1 dt
where the magnetic flux is given by
~=HB'ds (1.29)
It is also important to notice that the differential formulation of Faraday's

law is independent of the motion of the material through the field. This
enables us to transfer from a moving reference frame to a stationary one and
vice versa.
Finally, we recall the two relationships known as Green's theorem
HI (VU' vv + UV V) dv = ~ UVV' ds
2
(1.30)
HI (uv v - vv U) dv = ~ (UVV - VVU) . ds

2 2
(1.31)
which are particularly helpful in the formulation of the finite-element and

boundary-element methods.
1.2 SCALAR AND VECTOR POTENTIALS

One way of finding field distributions is to solve Maxwell's equations directly
as a system of first-order differential equations in terms of the appropriate
field quantities. In most, although not all, important cases, however, it will
be more economical to convert the first-order equations into second-order
differential equations involving scalar or vector potentials. Any vector field
F is uniquely defined if its sources are given as
divF = w (1.32)
and
cur/F =d (1.33)
where the source distribution w is a scalar field and the source distribution
d is a vector field. The total field F in the presence of both types of sources
can be obtained by superposition
Scalar and Vector Potentials 5
F = - grad t/> + curl A (1.34)

where the two components are independent of each other since
divcurlA =0 (1.35)
and
curl grad t/> =0 (1.36)
so that eqns (1.32) and (1.33) are satisfied through the following second-order
differential equations
div grad t/> = - w (1.37)
and
curl curl A =d (1.38)
It is, therefore, convenient to divide the field F into a gradient field and a
solenoidal field corresponding to the two types of sources, especially because
such sources are often easily distinguishable in physical problems.
In electrostatics we introduce the electric scalar potential in terms of the
electric field intensity as
E = -gradt/> (1.39)
or using the nabla notation
E = - Vt/> (1040)
Substituting eqn (lAO) into eqn (1.3) with the aid of eqn (1.6) yields
V· (EVt/» = -p (1.41)
and if E is constant within a region
V 2t/> = _.': (1.42)

E
which is a second-order differential equation, known as the Poisson 's

equation.
In magnetostatics we often use the vector potential A defined in terms of
the magnetic flux density B as
B=VxA (1.43)
With the help of eqns (1.9) and (1.7) we find
Vx (~V x AJ = J (1.44)
In conductors it is likely that 11 = 110' so that

V x V x A = VV· A - V 2A = 110 J (1045)
For convenience we are allowed to put

V·A =0 (1.46)
because there is a free choice in selecting the divergence sources of A, since
only the curl sources have been defined in eqn (1.43). The condition imposed
on the divergence of the vector potential is called a gauge condition, and the
particular choice of eqn (1.46) is known as the Coulomb gauge. Hence if we
choose zero divergence
(1.47)
In a rectangular coordinate system eqn (1.47) separates into three equations
V Ax = - ~oJx
2
(1.48)
V A y = - ~oJy
2
(1.49)
V 2A z = - ~oJz (1.50)
On the other hand, if the magnetostatic field is not required within the
regions of conduction current, eqn (1.9) simplifies to
vxH =0 (1.51)
and we have an alternative formulation available in terms of H. Thus we can
write
(1.52)
where <Pm is a magnetic scalar potential. This leads to the following equation
(1.53)
which when expanded yields
~V2lf>m + V~ . Vlf>m = 0 (1.54)
The formulation in terms of the magnetic scalar potential has some advantages
as it reduces the number of unknown functions which have to be solved from
three to one. In two-dimensional cases this does not matter as the vector
potential typically has only one component (in the direction of current flow),
but in three-dimensional fields savings in computing effort may be con-
siderable. Either formulation is complicated by the fact that permeability is
not only a function of position but depends also on field strength. Moreover,
many practical magnetic materials are anisotropic, and so the permeability has
to be treated as a tensor.
Generally, in time varying fields, the electric and magnetic fields are
coupled, as demonstrated by eqns (1.1) and (1.2). Both sources p and J are
present and are linked through the equation of continuity (1.5). Similarly the
electric scalar potentiallf> and the magnetic vector potential A can be combined
using the Lorentz gauge
1 olf>
V·A+--=O (1.55)
2
c ot
Scalar and Vector Potentials 7
where c is the velocity of light. The Lorentz gauge simplifies the relationship
between the potentials and the sources, which may be shown to be
I a2A
V 2A - 2- -2= -J,LJ (1.56)
c at
and
2 I a2~ p
V ~ - c2 at 2 = -g (1.57)
where J.L and E have been taken as constants. Thus A depends on J only,
whereas ~ depends on p only. The electric and magnetic fields are obtained
from the potentials by the relationships
E=
aA
---V~ (1.58)
at
and
B=VxA (1.59)
Other conditions on the potentials are possible through the use of the
general gauge transformation
A' = A + V", (1.60)
~' = ~ _ a", (1.61)

at
with a subsidiary condition
V 2 ",
a =0
I 2 ",
(1.62)
- -
c2 -2
at
Thus '" obeys the wave equation and the electric and magnetic fields of eqns
(1.58) and (1.59) are not affected by this transformation. Examples of possible
useful transformations include choosing ~' to be zero or making A' solenoidal
irrespective of the sources. A comprehensive survey of various possible
formulations may be found in the papers by Hammond [1.21] and Carpenter
[1. 7].
Finally, we turn our attention again to sinusoidally varying fields. It is often
convenient to separate the imposed current density from the conduction
current by writing
J = oE + L (1.63)
Using eqns (1.12), (1.13), (1.6) and (1.7) we can now write
Vx H = (0 + jroE)E + L (1.64)
V x E = -jroJ,LH (1.65)
Moreover,
D=VxA (1.66)
and
E= -jroA - V~ (1.67)
Taking the curl of eqn (1.66) in combination with eqns (1.64) and (1.7) yields
vx V x A = ~ (J + jroe)E + ~Js (1.68)
and thus
VV' A - V 2A = -jro~(J + jroe)A - ~(J + jroe)V~ + ~Js (1.69)
The Lorentz gauge may now be imposed in the form
V' A = -~(J + jroe)~ (1.70)
and with the aid of the following substitution
k 2 = jro~ (J + jroe) (1.71)
we arrive at two non-homogeneous Helmholtz equations linking current and
charge sources with the vector potential and the scalar potential, respectively,
as
2
V A - k A
2
= -~Js (1.72)
V2~ _ k2~ =_g (1.73)

e
The choice of an appropriate gauge for the vector potential is an important
matter in most computational schemes, and in particular in the finite-element
formulations, where some gauges lead to equations which may become badly
conditioned at very low frequencies and some loss in accuracy has also been
reported at high frequencies. For a comprehensive discussion of magnetic
vector formulations and gauge transformations in finite-element methods the
reader is referred to reference [1.5], pp.365-379.
1.3 CLASSIFICATION OF FIELDS

The following provides a list of the most common types of equations which
are frequently used to describe particular types of fields under static and time
varying conditions. Most chapters in this book describe ways of solving such
equations under particular conditions. Every equation has its scalar form and
a vector form, because they may be applied to a scalar potential or a vector
potential.
Laplace's equation:
(1. 74)
(1.75)
Boundary Conditions 9
Poisson's equation:
(1.76)
(1.77)
Helmholtz equation:
V2ep + k 2ep =f (1.78)
V 2A + k 2A = F (1.79)
Diffusion equation:
V2ep = a aep (1.80)
at
V2A = a aA (1.81)
at
Wave equation:
2
V2ep _ ~2 a ep2 =f (1.82)
at
V2A _ ~2a2A = F (1.83)
at 2
The above relationships fall into various classes of partial differential
equations. Thus the equation may be
(I) elliptical (eqns (1.74), (1.75), (1.76), (1.77), (1.78) and (1.79»),
(2) parabolic (eqns (1.80) and (1.81»),
(3) hyperbolic (eqns (1.82) and (1.83»).
1.4 BOUNDARY CONDITIONS

Two requirements are implied by the term boundary conditions. First, it is
necessary to contain the region in which a second-order partial differential
equation is to be solved by specifying on the boundary of the region sufficient
information to make the solution unique. Essentially, it is a matter of con-
veying to the interior solution the effect of any field sources outside the region
of interest.
Secondly, when two sub-regions having different magnetic or electric
properties meet, it is necessary to be able to link the solutions on both
sides of this interface. Although included under the generic name 'boun-
dary conditions', the latter are perhaps more appropriately called interface
conditions.
Boundary conditions at an interface

We shall derive the interface conditions at a magnetic boundary (and, by
analogy, a dielectric boundary), where in the more general case, two materials
of different relative permeability ~rl and ~r2 meet. Applying Gauss's theorem
to the surface of a coin-shaped volume that just squeezes a small piece of the
interface os between its two circular faces, we have
, B . ds = B2 • os - B( . os = 0 (1.84)
where the contribution of the ribbon edge of the coin-shaped volume is

negligible, and we are assuming that the positive direction of the magnetic field
is from region 1 to region 2. Thus
(1.85)
where the suffix n denotes the component of B normal to the interface.
In the analogous case of a dielectric interface, because of the possibility of
a free surface charge of density q (Cm- 2), the normal component of the
electric flux density is discontinuous and given by
(1.86)
An application of Ampere's law to a small rectangular path that just
squeezes the interface between its two longer sides Of yields
(1.87)
Whilst we can state that the tangential component of H is usually continuous
across the boundary, Le. H I2 = H/I' the tangential component is actually a
two-dimensional vector. Not only this, but if there is a surface current density
K (A m -I) on the boundary, eqn (1.87) can be usefully generalized as
(H 2 - HI) x n =K (1.88)
where, because K and the discontinuity in HI must be orthogonal, the vector
product with the unit vector n gives the direction of K.
In a similar way we can write, for the electric field E,
(1.89)
Considering again the magnetic interface, if a flux line in region 1 enters the
surface at an angle 91 to the normal and exits into region 2 at an angle 92,
then eqns (1.85) and (1.88) can be written
B 2 cos 92 = B I cos 91 (1.90)
H 2 sin 92 = HI sin 9 1 (1.91)
Dividing eqn (1.91) by eqn (1.90), and using eqn (1.7), gives
tan 9 I!o I!r2
- - 2= - - = -
I!r2
(1.92)
tan 91 I!Ol!rl I!r]
Suppose now that region 1 is very highly permeable and region 2 is air. The
right-hand side of eqn (1.92) tends to zero and it therefore appears that 92
must be very close to 0 0 , i.e. the flux enters or leaves a very highly permeable
surface at right angles. This is indeed so under most conditions, but there are
Boundary Conditions 11
circumstances where 92 can be considerably greater than zero even with very
high values of Il, I.
If, for example, we consider the field inside a cylindrical tube of high
permeability situated in a uniform field transverse to its axis, the field in the
screened region is always parallel to the applied field and so cannot leave
and enter the curved interior surface everywhere at right angles. The reason
is the dominant tangential field within the tube wall which may cause 91 to
be very close to 90 0 over large portions of the surface. If 1l,1 = 1000 and
91 = 89.9 0 , for example, eqn (1.92) yields 92 = 29.8 0 , Le. considerably
greater than would normally be expected on the air side of a highly permeable
surface.
Surface polarity
Although the source of a magnetic field is electric current, it is sometimes
convenient to introduce magnetic poles or polarity density qm to represent the
effect of a magnetic surface [1.40; 1.5]. qm is analogous to the bound electric
charge density that can be used to model a dielectric surface. Note that q in
eqn (1.86) is the free electric charge density which has no counterpart in
magnetism; the bound charge is hidden in the definition of electric flux density
D. If qm is the local value of surface magnetic polarity, then, by Gauss's
theorem,
(1.93)
where H~ is the normal component of the magnetic field due to qm. If the
normal component of the applied field due to all other sources is H m eqns
(1.85) and (1.7) yield
lloll,2(Hn - H~) = 1l01l,I(Hn + H~)
or
H' = 1l,2 - 1l,1 H (1.94)
n 1l,2 + 1l,1 n
from which qm follows in terms of H n. Region 1 is assumed here to be the

region of lower permeability, usually air, and it may be more convenient to
express qm in terms of the total normal component H n 1 = H n + H~. Thus
from eqns (1.93) and (1.94)
qm = (1l,2 - 1l,1) ~ H nl (1.95)

1l,2
It is important to take care with the use of eqn (1.95) because it is an alternative
way of representing the magnetized surface and restricts us to using the H field
or the scalar potential related directly to it (eqn 1.52).
External boundary conditions

Some problems have open boundaries that extend to infinity in a mathe-
matical sense. Analytically this is not usually difficult, but to deal with such
problems numerically may require special techniques. More usually a prob-

lem will have a natural physical boundary on which one of the components
of the magnetic field may be zero or can be assumed to be a well defined
function of position. We note that, from eqn (1.95), a surface layer of
magnetic poles can cause a discontinuity in the normal component of H, and,
from eqn (1.88), a surface current can set up a discontinuity in the tangential
component of H.
We are thus able to convey information about magnetic fields and field
sources external to the region of interest via specification of either HI or
H n over parts or the whole of the selected boundary surface. However,
specification of H n is not sufficient without imposing additionally the satis-
faction of Ampere's law for the volume concerned. The three-dimensional
generalization of Ampere's law is [1.38]:
Ifs (n x H) dS = JHcurlHdV = JHJdV

v v
(1.96)
In a two-dimensional problem with the current flow perpendicular to the plane

of the region eqn (1.96) reduces to
(1.97)
In terms of the scalar magnetic potential 4>m (eqn 1.52), HI corresponds to

specification of 4>m. This potential specified boundary condition is known
as a Dirichlet condition. Alternatively, H n leads to iJ4>m/iJn, the Neumann
condition.
In terms of the magnetic vector potential A (eqn 1.43), we require
n x curl A or n x A respectively, which reduce, in the two-dimensional case
where the only component of A (A p say) is perpendicular to the two-
dimensional plane, to the Dirichlet specification A p (for H n) and the
Neumann specification iJAp/iJn (for HI).
1.5 MAGNETIC MATERIALS

Equation (1.7) gave the linear relationship between Band H. More precisely
we have
B = J.1 oH +M (1.98)
where M is the magnetization vector which is the magnetic moment per unit
volume. Only in linear isotropic materials is M both parallel to and propor-
tional to H, Le.
M = J.10'XmH (1.99)
where 'Xm is a constant known as the magnetic susceptibility. Equation (1.98)
then becomes
(1.100)
Magnetic Materials 13
where the relative permeability

J.1, = 1 + Xm (1.101)
Although eqn (1.100) is an approximation for many magnetic materials, and
it is often the practice to allow J.1, to vary as a function of H, it does allow
us to classify magnetic materials into three categories:
Diamagnetic Xm is a very small negative number (J.1, < 1),

Paramagnetic Xm is a very small positive number,
Ferromagnetic Xm is a large positive number (J.1, ~ 1).
Diamagnetism and paramagnetism are very small effects, with Xm typically of

the order of 10- 5, and have to do with the orbital motion and spin of the
electrons within the material, but ferromagnetism is very strong and a vital
feature in a wide range of power apparatus. Ferromagnetic materials are made
up of domains which are already magnetized in the sense that each domain
contains aligned magnetic dipoles formed by electron spin. When the sample
of material is in an unmagnetized state the domains are so aligned that their
net magnetization cancels. For example, in an iron crystal there are six direc-
tions of 'easy' magnetization: the positive and negative directions along three
orthogonal axes. If we assume, for the purpose of the argument only, that
all domains are the same size, then in each of the 'easy' directions there will
be as many domains pointing in the positive as in the negative directions. Now
if a small magnetic field is gradually applied to the unmagnetized crystal nearly
parallel to one of the easy directions, the domains with polarity in the same
direction, Le. assisting the applied field external to themselves, begin to grow
at the expense of adjacent parallel domains with opposite polarity, by virtue
of small movements of the domain walls. This process is virtually loss free
if it takes place slowly.
Let us move now to consideration of a larger sample made up of many
crystals or crystal fragments, each containing many domains, in the form of
a ring on which there is a uniformly wound toroidal magnetizing winding
producing a uniform applied field H parallel to the internal axis of the ring
cross-section. This configuration gets over the problem of possible differences
between the applied and the local field because there is no surface magnetic
polarity present. Thus, with a simple search coil to measure the corresponding
flux density B, a curve of B against H, starting with an unmagnetized ring,
can be drawn as shown in Fig. 1.1. The previously described process with a
low value of H applied, takes us to point PI' If this field is removed the
operating point drops back to the origin because the process is reversible. To
reach the higher point P 2, the domains already aligned with the field direc-
tion continue to grow, other opposing domains completely reverse, and those
at other angles reverse polarity if this brings them more nearly into line with
the applied field. All this activity involves extensive domain wall movement
and the expenditure of energy. As a result when the field is reduced to zero
the flux density only drops to the point represented by Q2' and if the applied
RI3
Fig. 1.1 Hysteresis loops for a ferromagnetic material
field is then cycled between equal negative and positive maximum values the
complete loop P2 P i. is repeatedly traversed. This is known as a hysteresis
loop, derived from the Greek word meaning 'to lag', and it can be shown that
the expenditure of energy per cycle is given by the area of the loop. Alter-
natively, rising up the initial magnetization curve (also called the normal
magnetization curve) to the higher point P 3 takes the material into magnetic
saturation whereby almost complete alignment of the domains in the field
direction has taken place. For a full discussion of the processes involved see
Bozorth [1.6] or Chikazumi [1.l2].
Symmetrical hysteresis loops, know as major loops, can be formed from
any points such as P 2 or P 3 • The value of flux density at which H = 0 on any
loop is known as the remanent density B, and the value of H at which B = 0
is the coercive force He. The so-called permanent magnetic materials operate
in the second quadrant of Fig. 1.1 and require a very high coercive force, or
coercivity, to prevent demagnetization during operation.
Hysteresis loops do not have to be symmetrical. An important example
occurs if we are on a major loop such as that represented by P 3Pf, in Fig. 1.1
but stop at point R 3 and begin to reduce H again but only by a small amount
to point S3' By causing H to oscillate by t1H produces the minor hysteresis
loop shown. The mean slope of this loop, Le. t1B/t1H, is known as the
incremental permeability ~i' Minor loops can also be generated from points
on the initial magnetization curve. The incremental permeability at a given
point decreases with dH and the limiting value of ~i as t1H tends to zero is
the reversible permeability ~,ev' It is necessary to consider the incremental
or reversible permeability in situations where a small a.c. perturbation exists
in the presence of a much larger d.c. field, because, for example, the eddy-
current loss can be much greater than it would be if only the absolute per-
meability (~o~,) or differential permeability (dB/dH) is taken into account.
Energy Storage and Forces in Magnetic Field Systems 15
1.6 ENERGY STORAGE AND FORCES IN MAGNETIC

FIELD SYSTEMS
In the magnetic field model, the energy associated with the system is distri-
buted throughout the space occupied by the field. Assuming no losses, energy
supplied per unit volume (the energy density) when increasing the flux density
from zero to B is
(1.102)
The total magnetic energy stored is
(1.103)
If, for simplicity, we assume a uniform distribution of field in the region

concerned of volume v, then
B
Wm=vJHdB (1.104)
o
If, in addition, we have a linear magnetic material
1 2
Wm = v-B (1.105)
211
Clearly the higher 11, the lower the total energy stored. For the same flux
density, the energy stored in a given volume of air is orders of magnitude
higher than in the same volume of iron.
For a singly excited system, eqn (1.104) can be written in the form
B ~
Wm = J (H x length) (area x dB) = J mmj(cP) dcP (1.106)

o 0
where cP is the total flux passing through the region. But mmj = ni where n
is the number of turns on the excitation winding and cP = 'AJn where A. is the
total flux linkage. Thus
i
Wm = Ji(A.) dA. (1.107)

o
Note that the current is a function of the flux linkages and the mmj is a
function of the flux; their relations depend on the geometry of the magnetic
circuit and windings, as well as the magnetic properties of the core material.
Equations (1.104), (1.106) and (1.107) may be graphically interpreted as the
area shown in Fig. 1.2 labelled as energy. The other area, labelled as coenergy,
can be expressed as
H or i or mm!
Fig. 1.2 Energy and coenergy
Io BdH = I1..(;) di = I (j>(mmf) dmmf

H i mmJ
W,;. =V (1.108)
0 0
For a linear system in which Band H or A. and i or (j> and mmf are pro-
portional, the energy and coenergy are equal. However, for a nonlinear system
the energy and coenergy will differ but their sum for a singly excited system
is given by
W,;. + Wm = vBH = 'Ai = mmf(j> (1.109)
Consider a simple electromechanical system where movement is allowed in
one direction only, say x, and electrical and mechanical losses are neglected.
An energy balance equation may be written
electrical energy input = mechanical energy output
+ increase in stored energy
In the simple situation of a change in x at constant A. there is no emf induced
and therefore no electrical input so that the force is given by
F-- - (awmJ
ax ')..=constant
(1.110)
Consider now Fig. 1.3 which shows a change of Wm at constant current

and the associated electrical input due to the changing flux. We can now
write
F = i [a1..J - [awmJ (1.111)

ax i=constant ax i=constanl
or, in terms of the coenergy
Energe Storage and Forces in Magnetic Field Systems 17
A
x=dx
A2
x=O
A1
Fig. 1.3 Change of energy at constant current
F = [a(iA.)] _(awmJ = (aw,;,] (1.112)

ax ;=consla,,1 ax ;=consla,,1 ax ;=consla,,1
It is interesting to notice the change in sign between eqns (1.110) and (1.112).
If the relationship between A. and i is a straight line the magnitude of the two
expressions for the force is the same but the sign is reversed because of the
electrical input.
Calculations of forces and torques based on the energy method described
briefly above are usually quite convenient. They involve volume integration
of energy or coenergy and such routines are usually provided by post-
processors of commercial software. However, at least two field solutions
are required in order to determine the space derivatives. For improved
accuracy several solutions may be necessary and thus computation times may
be long.
Another approach, known as the stress method, looks at tensile and com-
pressive stresses along iron surfaces and relies on surface integration of
appropriate expressions. We have already seen that it is possible to represent
surface magnetization in terms of a distributed magnetic polarity density qm'
If, in eqn(1.94), ~'I = I, i.e. region 1 is air, then from eqn(1.93)
~, - 1
qm = 2~o--lH"
~r +
(1.113)
Combining eqns (1.95) and (1.113) we also obtain

IJ.r + 1
H"=-2- H ,,1 (1.114)
IJ.r
Now, since the force on a single pole Q is given by QHa, where H a is the
applied field, the force per unit surface area on qm is given by the stress
components
Fn = qmHn and FI = qmHI (1.115)

Thus, after substitution, we have
IJ., - 1 2 1 2( -2) (1.116)

Fn = 2IJ.o-- H n = -2IJ.o H nl 1 - IJ.,
IJ., +1
IJ., - 1 (-I)
F I = 2IJ.o-- H nH I = IJ.OHnIHII 1 - IJ., (1.117)
IJ., + 1
since the contribution of qm to the local tangential field is zero, and so the
applied tangential component of field HI is equal to the total HI I' In the
common case where IJ., ~ 1, H tl --+ 0 and Fn --+ IJ.oH~ /2 and FI --+ O. These
stresses are not unique, but since the surface polarization correctly reproduces
the magnetic field conditions at all points outside the iron body, the total force
on the iron obtained by surface integration is unique.
The full equations for Fn and FI only apply to the surface of a homo-
geneous iron region with a uniform value of IJ.,: otherwise there is a volume
distribution of polarity present. It would be useful to have more general
expressions for Fn and FI that could be applied to any closed surface in air
that surrounds a composite device consisting of nonlinear steel, current
carrying conductors etc., in the confidence that the resultant force on whatever
is within the enclosed region is correctly obtained by surface integration of
the two stress components. Such expressions are Maxwell's stresses in air
(vacuum) [1.16]:
Fn = t IJ.o (H; - Hf) (1.118)
F I = IJ.oHnHI (1.119)
where F n is directed outwards from the surface. The suffix 1 is no longer
required to distinguish the region exterior to the iron from the interior: H n
and HI now represent the components of the magnetic field at the chosen,
and, to some extent, arbitrary surface. Whereas the earlier expressions can
be thought of as stresses acting on the iron surface, the much more general
Maxwell forms represent stresses transmitted through the chosen computation
surface by the magnetic field.
1.7 EDDY-CURRENT SKIN EFFECT

If a very large flat conducting surface on the plane y = 0, say, is subjected
to a uniform alternating field parallel to the surface, Hz = H s cos rot, then
the solution of the complex one-dimensional differential equation (Helmholtz
equation) for Hz inside the conducting slab is
H = H s e- ay
-z (1.120)
where a = (l + j)/o and 0 = (roIJ.o12)-1/2. The corresponding current density
distribution is given by
Eddy-Current Skin Effect 19
J
"-X
= dB
- z = -aH e-
dy S
ay
= -J e- ay S
(1.121)
and the eddy-current loss per unit surface area of the slab is
(1.122)
But, using Ampere's law, H s = I, the peak current per unit width. Thus the
loss can be written in the form
[~r a 0 = I;ms R '

1
Pe = (1.123)
where R' = (ao) -1 is the resistance of unit length of a layer of thickness 0

and unit width. For this reason the parameter 0 is known as the eddy-current
skin depth because the correct value of loss is obtained if the current is
assumed to flow with uniform density in a layer of depth o.
In practice, the magnetic field and current density fall exponentially in
magnitude away from the surface. Not only this, but there is a phase change
with increasing depth. For example, from eqn (1.121), the instantaneous value
of the current density is given by
Jx .= Re (,Lei<ot) = -Jse-yl15cos [rot - ~J (1.124)
so that, when y = no, Jx is in anti-phase with its surface value (-Js)' This is
typical of a diffusion process. At any depth, the local peak value does not
occur until y/roo seconds after it occurs on the surface, and, from one point
of view, it is as if a damped wave of current (or flux) is moving into the mate-
rial at velocity roo.
The relative smallness of the total penetration is particularly important. If
we regard anything less than 50/0 of the surface density as negligible, then the
effective penetration is given by 30. For copper at 50 Hz, 0 is typically about
9 mm so that the penetration, on this criterion, is 27 mm. But 0 is inversely
proportional to the square root of frequency so that, at 500 kHz, the effective
penetration has reduced to 0.27 mm. The same sort of reduction occurs even
at power frequencies if the material has a high relative permeability. The result
is that this simple one-dimensional approach can be used to model, at least
approximately, the surface of conductors of any shape if the major dimensions
of the conductor are much greater than 0, because locally the surface effect
is almost one-dimensional. We only require to know the distribution of the
surface field H s • In numerical finite-difference or finite-element solutions this
can be particularly important because the alternative will be the need to
discretise the grid or mesh sufficiently finely to follow the skin phenomenon
correctly.
The simpler numerical technique is to render the conducting region imper-
meable to flux by setting 1-1, = 0 and therefore excluding this region from
the computation by specifying H n = 0 on its surface. A practical example,

although most often solved analytically, is the wall of a waveguide.
In situations where the skin effect is rather less pronounced, but the
penetration nevertheless still small, the conducting surface can be modelled as
a surface impedance. This is the equivalent of a local linear one-dimensional
solution because the surface impedance is defined as
(1.125)
where Es is the peak value of the surface electric field. From Ohm's law,
eqn(1.8), and eqn(1.121),
aH
E J
= ( 1 + J.) 00
H
s s s
s =~ = --;;- (1.126)
Thus
1+j
Z =-- (1.127)
s 00
The above situation is one in which the eddy currents are limited by their
own field. The currents are said to be inductance limited. There is an opposite
extreme when the currents are resistance limited owing to lack of space. This
is precisely what is required in a transformer lamination, for example. The
full linear solution for the eddy-current loss per unit surface area of a plate
or lamination of thickness 2b is [1.39]
P = H; sinh 1 - sin 1 (1.128)

e 00 cosh 1 + cos 1
where 1 = 2b/o. It is clear that as 1 tends to infinity, the loss tends to
H;/oo, Le. twice the loss per unit surface area of a massive slab, because we
have two surface effects. However, when 1 <lli 1, the loss tends to
3
P' = H; 1 (1.129)
e 00 6
which is the result that would have been obtained by ignoring the reaction field
of the eddy currents altogether.
The condition for this resistance-limited regime is that the thickness of the
conductor is less than o. If such a conductor is present in a field problem it
can be modelled as having zero conductivity, Le. no eddy currents, as far as
the external system is concerned. Clearly, if it is a load carrier, an imposed
uniform current density must be specified.
Geometrical Structure of Electromagnetic Fields and Differential Forms 21
1.8 GEOMETRICAL STRUCTURE OF ELECTROMAGNETIC

FIELDS AND DIFFERENTIAL FORMS
Vector algebra provides a familiar notation for describing electrom~gnetic
relations and is used throughout this book. However, there is much to be said
for using differential forms instead of vectors, because these forms underline
the geometrical structure of the equations. Whereas the vector calculus cannot
extend to four dimensions, both three- and four-dimensional problems can
be readily treated by differential forms. Other advantages of such treatment
include the unification of static and dynamic fields, the clarification of the
difference between gradient (polar) and flux (axial) vectors and the replace-
ment of the operators grad, div and curl by a single exterior derivative. In this
section we shall introduce very briefly this powerful notation, in which
physical and geometrical information is combined. For full treatment the
reader is referred to specialist literature, in particular works by Hammond and
Baldomir, for example references [1.2; 1.3; 1.4; and 1.22].
In terms of differential forms, Maxwell's equations can be expressed as
[1.1]
dD=p, dH=J+aID, dB=O, dE= -alB (1.130)
together with the constitutive relationships
D= &• E, B= ~ •H (1.131)
The variables are how known as forms: a O-form is a scalar, a I-form has a
short length associated with it, a 2-form an area and a 3-form a volume. For
example, the 2-form B is equivalent to B . os in vector notation. Hence we·
can write
cP Q-form scalar potential

A I-form magnetic vector potential
E I-form electric field strength
H I-form magnetic field strength
D 2-form electric flux density
B 2-form magnetic flux density
J 2-form electric current density
p 3-form electric charge density
In the notation of eqns (1.130) and (1.131), d is the exterior differential

operator (which corresponds to divergence and curl in vector analysis), l is a
the differential operator with respect to time and • denotes the Hodge opera-
tor linking a differential 2-form (flux) having area with a differential I-form
having length. This linkage defines a volume element involving the product
of two field quantities. The products of D and E and of Band H have the
dimensions of energy, and the constitutive equations therefore define the local
energy density associated with a local volume.
*
r--------------~
O· form
I 1 - form
1*
I
I
2 - form
3 - form
4 - form
Fig. 1.4 Maxwell's equations in differential forms
It is interesting to note that the double exterior differentiation is always

zero, Le. d· dct> = 0, d· dE = 0, d· dD = O. This is a general property of
differential forms and is known as Poincare's lemma. It corresponds to
curl grad ct> = 0 and diu curlE = 0 in vector algebra. Notice also that a I-form
is always associated with an element of length and a 2-form with an element
of area. This avoids the confusion which often arises between axial and polar
vectors in vector analysis.
Equations (1.130) and (1.131) can be simplified by introducing the
potentials A and ct>, together with a gauge condition between them. We obtain
OA = -IJ,*J (1.132)
I
Oct> = - -e * p (1.133)
where the Lorentz gauge has been used. This gauge is given by
oA + lJ,eotct> = 0 (1.134)
Here the operator 0 is the co-derivative of the exterior differential operator
d and is defined as 0 = *d* (-1 )p(n-p+l)+S, where n is the dimension of
space, p is the order of the differential form and s is the number of negative
terms in the metric. The Laplace-Beltrami operator 0 is defined as
o = do + Od. In three dimensions, OA is equivalent to grad diu - curl curl in
vector notation, and Oct> is equivalent to diu grad.
Maxwell's equations can be represented diagrammatically as in Fig. 1.4
[1.1]. Each level of the diagram corresponds to a different form of a particular
order. Horizontal arrows with their signs indicate the time derivative at, and
the vertical arrows indicate the exterior derivative d. Convergence of two
arrows at a node implies summation. The dotted lines show the constitutive
Geometrical Structure of Electromagnetic Fields and Differential Forms 23
relations. Quantities, obtained by summing terms in which both d and at are

involved, are shown in squares. The scalar potential q, is a O-form and the
vector potential A is a I-form. The introduction of the potentials makes the
right-hand side of the diagram complete.
The scalar and vector potentials are related to the field vectors via the
following relationships
B = dA (1.135)
and
(1.136)
Figure 1.4 displays a particular type of symmetry, where two blocks are
displaced with respect to each other, as regards the order of the differential
forms involved. The two blocks are linked through the constitutive relation-
ships, eqn (1.131). The symmetry between the blocks allows the correspon-
dences [1.1]
dp = 0 =:: dB = 0 (1.137)
dD=p=::dA=B (1.138)
dB = J + atD =:: - dq, = E + atA (1.139)
dJ + atp = 0 =:: dE + atB = 0 (1.140)
in which each of Maxwell's equations is paired with a relationship contained
in the other block. For example, eqn (1.137) states that the electric charge
density p is locally conserved, like the flux associated with the magnetic
induction B. Equation (1.138) parallels Gauss's law to the existence of vector
potential. Equations (1.139) and (1.140) embody Ampere's law and Faraday's
law, as well as the local conservation of sources.
Electrostatic and magnetostatic fields may be retrieved from Fig. 1.4 by
prolonging the dotted paths joining the fields with their duals as far as
possible, on either side, without introducing time-varying quantities. The
relevant equations are found for electrostatics as
(1.141)
and for magnetostatics

(1.142)
Both pairs of relationships lead to Poisson's equations. The corresponding
diagrams are depicted in Fig. 1.5.
The product of E and D (and H and B) in the notation of differential forms
is a 3-form and is, therefore, associated with a small volume of geometrical
space. Hence the division of space into 'cells' implies the division of the energy.
Another 3-form is associated with the 'assembly work' of, say, charges
described by the product of q, and p. The integrals of the field energy and
Electrostatics:
Magnetostatics:
Fig. I.S Paths for electrostatics and magnetostatics
the assembly work are equal for the system, but locally the product </>p can
be positive, negative or zero, whereas the product ED is non-negative. We
can describe the field in two different ways. The first starts with the I-form
(E or H) and its associated increments of potential difference, whereas the
second begins with the 2-form (D or B) which gives a flux through an area.
Application of the Hodge operator and identification of relevant area or line
segment produce a volume and an associated energy. These elementary
volumes are then joined in parallel or series to form 'slices' or 'tubes', which
are then connected in series or parallel to fill the entire space of the system.
The implementation involves simple multiplication, followed by addition in
series and parallel; it does not require the solution of simultaneous equations.
The technique is known as the method of tubes and slices (1.22; 1.43] and is
briefly described in the next section.
1.9 THE METHOD OF TUBES AND SLICES

The method of tubes and slices is based on a dual energy formulation. Founda-
tions of this approach may be traced back to Maxwell, who in his famous
treatise on electricity and magnetism [1.30], describes a variational method
applied to the calculation of resistance of conductors of varying cross-section.
The method relies on subdivision of the conductor into slices formed by
equipotential surfaces and tubes separated by very thin insulating sheets. The
two calculations yield lower and upper bounds of the resistance, respectively.
The approach is applicable to other types of vector fields as demonstrated
by Hammond [1.20; 1.22]. In electrostatics Maxwell's method of 'slices' is
identical to the application of the variational principle
«p - V . D), o</» = «( p + V' (eV</» l. o</» = 0 (1.143)
where the brackets <> indicate integration through the region of interest, and
his method of 'tubes' is identical to the application of
«E + V</». oD) = ((~D + V</>J. OD) = 0 (1.144)

The Method of Tubes and Slices 25
The two functionals are equivalent to introducing additional fictitious

sources, of divergence or curl type respectively. In capacitance problems where
the energy can be written as ! fj> 2C or ! Q2/C, upper bounds of the capaci-
tance are produced by the variational statement of equation I (slices) and
lower bounds result from the use of equation 2 (tubes). Thus any potential
map produces an upper bound of capacitance and any flux map produces a
lower bound.
In magnetostatics the equilibrium conditions can be described by two
variational principles
( (V x H - J' ), oA) =0 (1.145)
and
«(VxA-B),oH)=O (1.146)
where J' is the assigned current density. The first variational principle implies
that B = V x A, so that V . B = 0 and thus there are no divergence sources for
the magnetic field. However, the expression V x H - J' allows a small varia-
tion in V x H from its correct value, ·so that the variation allows a small
additional distribution of curl sources. The product of this small fictitious
current multiplied by the small variation of A gives an energy variation of the
second order of small quantities which can be put to zero.
The second variational principle assumes that V x H = J' , so that the curl
sources of the magnetic field are correct. However, the expression V x A - B
allows a small variation in the divergence sources. The product of this small
polarity distribution multiplied by the small variation of H gives an energy.
variation of the second order of small quantities which can be put to zero.
The field energy can be expressed either in terms of the field vectors Hand
B by
U=! (B,H) (1.147)
or in terms of the interaction of the current sources with the vector potential
A by
U=!(J',A)+![I',A] (1.148)
where I' is the assigned line density of current on the surface, and the brackets
[] represent integration over the closed boundary surface. I' is related to J'
so as to make the total current in the system zero. This isolates the system
and gives it a unique energy.
The first variational principle is applied to the energy in terms of A and
B = V x A by writing
oU(A) =O(J"A) + [I',A] -~(B,~)J =0 (1.149)
The second variation is therefore negative

(1.150)
The second variational principle is applied to the energy in terms of H by

writing
oU(H) =o(-!<H,IlH») =0 (LIS!}
Hence
(1.152)
For simplicity Il has been assumed constant and this gives the factor 1/2.
However, the method is applicable to permeabilities which are single-valued
functions of the field strength [1.20]. The second variations show the possi-
bility of obtaining both upper and lower bounds for the energy. The first
variational principle treats the field as a collection of tubes and the second
one as a collection of slices.
Similar arguments may be presented for a steady current flow and calcu-
lation of resistance, where the analogue of permittivity e or magnetic per-
meability Il is electric conductivity o. Extension of the method to time-varying
problems is possible and some interesting suggestions may be found in refer-
ences [1.31; 1.4; and 1.20].
The implementation of the method is relatively straightforward. For
example, in electrostatic problems, for a given field solution, the flux map and
the potential map may be represented using connections of simple parallel-
plate capacitors, as demonstrated in Fig. 1.6.
Tubes:
(a)
c.. = e X width (F/m)

I) mean length
The Method of Tubes and Slices 27
Slices
(b)
C ..
IJ
=E x mean width (
length
F/m
)
Fig. 1.6 Circuit representation of tubes and slices in electrostatics
These component capacitors are connected in parallel and in series, as shown

in Fig. 1.7, to form an equivalent circuit of the field map.
Tubes:
(a)
Slices:
(b)
Fig. 1.7 Series/parallel connections 0/ component capacitors
Neither of the two representations is exact owing to the approximations intro-

duced. First, the number of subdivisions will necessarily be finite. Secondly,
the flux or potential lines may not be in the correct position so that the
orthogonality of the two field maps is violated. This second consideration is
very important and leads to the following equation for tubes
n 1
c- = ~-=-m- (1.153)
i=I~_l_
j=1 [e~jJ
whereas for slices
1
C+ = --=m=------ (1.154)
~ 1
i=1 ~ reS.~
j=1 l/;JJ
where Sij and Ii) are dimensions of the component capacitors. Thus dual
bounds for the capacitance are established.
For the steady current flow problems the resistance may be calculated in
an analogous way. Physically, subdivision into tubes can be achieved by
inserting thin insulating sleeves between the tubes, which must always increase
the resistance unless the flow of current is undisturbed. Very thin sleeves in
the correct direction everywhere will have negligible effect, but if such sleeves
are not strictly in the direction of current flow they will increase the resis-
tance. The undisturbed resistance is therefore a minimum. Equally, the inser-
tion of infinitely conducting sheets for slices will reduce the resistance if they
Field Calculations using Tubes and Slices 29
disturb the flow. The undisturbed resistance is therefore a maximum. Thus

a division into tubes and slices enables one to calculate upper and lower
bounds for the unknown resistance.
Finally, in many electrical devices the magnetic circuit is designed in such
a way that very little mmjis absorbed in the iron core and attention is focused
on the shape and dimensions of the air-gap. An unsaturated iron surface may
be assumed to have a constant magnetic potential and thus becomes a slice.
At the same time the flux distribution may be described in terms of tubes.
These tubes terminate on iron surfaces. We can work in terms of permeance
which is the analogue of conductance in the electric circuit. However, for
problems outside a current region the calculation of inductance is reduced
to a calculation of permeance. Thus, for calculating inductance, equations
analogous to (1.153) and (1.154) may be used with I.l. substituted for E. We
now have a system of equations for calculating circuit parameters R, C or L
for many practical problems under static conditions.
Calculation of an internal inductance, on the other hand, involves working
in terms of energy instead of the circuit parameters. A possible approximation
technique is described for example in reference [1.43].
The simplicity of the final expressions (eqns (1.153) to (1.154)) is very
striking and should be contrasted with the complexity of other numerical
formulations, such as finite elements. Although the TAS method, like indeed
many other methods, usually calls for more than one calculation to achieve
the desired accuracy, this should hardly matter in view of this simplicity of
computation.
1.10 FIELD CALCULATIONS USING TUBES AND SLICES

The tubes and slices method is essentially a geometrical approach, based not
on solving equations but using sketching of fields as a means of finding the
solution. Full advantage can be taken of interactive graphics capabilities of
modern computers. An appropriate computer package called TAS is described
in reference [1.43]. In particular, the program has been found extremely useful
in teaching, as discussed in reference (1.17] and fully implemented in a recent
book [1.16].
In the TAS program the user is in'constant interaction with the field solution
provided in the form of approximate distributions of tubes and slices. These
distributions may be easily modified by moving appropriate construction
points and lines on the screen and the system responds almost instantly with
a pair of bounded solutions given by simple calculations of the type given by
eqns (1.153) and (1.154). The bounded values give confidence limits to the
solution, whereas the orthogonality of the two field patterns, or lack of it,
sets a criterion for further modifications.
In order to demonstrate the application of the method, let us estimate the
inductance per unit length of the coaxial line, with a strip centre conductor,
as shown in Fig. 1.8. Owing to symmetry only a quarter of the system needs
to be investigated (see Fig. 1.9). A possible set of 'construction lines' and
y ~~
-----.~
4 4 8
I' "I
-
X ..
I~
10 10
'"
"
Fig. 1.8 A rectangular coaxial line with a strip centre conductor
y
8
t slice boundary
6
tube slice
boundary boundary
4
0
0
slice boundary
2 4
tube boundary
6 8 10
-
X
Fig. 1.9 Computational model of the coaxial line
'quadrilaterals' is demonstrated in Fig. 1.10. Such quadrilaterals (each with a

diagonal line) are used to match a particular shape of system boundaries.
These construction lines and quadrilaterals are at the same time used to
generate tubes and slices of Fig. 1.11. The whole process is fully automated,
although the user may wish to change manually some distributions. An
improved subdivision may be achieved by adding more construction lines to
the basic distribution of Fig. 1.11 and repositioning them so that a more
orthogonal system is obtained. An example of such an improved distribution
is shown in Fig. 1.12. The results are summarized in Table 1.1 and compared
with a finite-element solution. It is interesting to note that although the
Field Calculations using Tubes and Slices 31
y
8~--------~
6
0-1-.,.-.,.-iil-';---,---¥-~-~+-----,
o 2 4 6 8 10
x
Fig. 1.10 Construction lines forming a set of quadrilaterals
II
,
• 4,----;r-----;r-----..,...----..,. ,.----";..=:.:..:;..---~
H: a.oo. opt••
, 1m IOtIl: Illit tM PNI1'M
Fig. 1.11 Distribution of tubes and slices
Table 1.1 Summary of results
Coarse Improved Finite

distribution distribution element
LHm- 1 0.9817 x J1 0.9968 x J1 0.9993 x J1
(per unit length)
Confidence limits ± 17.1070 ± 7.4070 ±0.7OJo
Actual error -1.76070 -0.25070
(against FE)
18
SI-r-I--L
AUlll'alJlI: 8.99611 J1
1 7.4;(
"-:
UlI: a.- apUIlill
H: a.-aptllllll
ION: Itllt till ~
Fig. 1.12 Improved distribution of tubes and slices
improved solution exhibits improved orthogonality, the average value of the

capacitance is hardly changed. Nevertheless, the smaller error band of the
improved solution gives additional confidence to the user.
The TAS program has a built-in finite-element module. This combination
of the two techniques has been found to be very useful. First, an approximate
field distribution is calculated by the tube/slice process and then the accuracy
is improved by means of a finite-element procedure. A much smoother distri-
bution is obtained as shown in Fig. 1.13. It should also be noted that the finite-
element method has been adapted to produce upper and lower bounds to the
unknown exact values.
The two methods complement each other. The construction lines and
quadrilaterals from the TAS calculations are now used as a basis of a FE mesh.
This mesh will typically be refined automatically to improve accuracy. The
values at all nodes are first calculated from the TAS solution so that a
relatively small number of iterations is required for the solution to converge.
It is also worth noting that the mesh is very regular and that most mesh lines
follow the direction of the field or the direction of equipotential surfaces.
Further information about this powerful combination of the two techniques
may be found in references [1.17] and [1.44].
1.11 PREPARATORY AND FURTHER READING

Before proceeding to the next chapters which discuss particular methods and
techniques of computational magnetics, the reader may have decided to study
further the fundamentals and background theory. If the short review provided
Preparatory and Further Reading 33
,
8 ~--r-,------,r----.-----,--------f 1-....;;....--"----1
7 t---t--I--J..._
uti:
....
tube -
a.. opu.
:
slice
H: a..opU.
h~a-...........o;-l-.l...:-l.....L..;~~4.l...L~5....L...-~*7 -l.,:--.l..:9:---:1~a 101M: JtWftpt
Fig. 1.13 Finite-element solution for the coaxial line
in this chapter has not been sufficient, we would like to make some recom-
mendations to aid such studies. A very comprehensive treatment of both
the theory and computational methods at the undergraduate level is offered
in the book by Hammond and Sykulski [1.16J, where the reader will find a
full derivation of most of the equations reviewed in this chapter, comple-
mented by thorough discussion of underlying physical processes and details
of the numerical implementation of methods.
For a more introductory treatment of the subject the reader may wish to
refer to the textbooks by Carter [1.10], Christopulos [1.13], Hayt [1.24] or
Lancaster [1.27]. It is also appropriate to mention some more advanced texts
by Carter [1.11], Dobbs [1.14], Harrington [1.23], Moon and Spencer [1.31],
Smythe [1.36] and Stratton [1.41]. A recent significant book on analytical and
numerical solution of electric and magnetic fields is by Binns, Lawrenson and
Trowbridge [1.5].
Eddy currents have always been of particular interest to both researchers
and practising engineers, and the following three books may be recommended
for full treatment in various aspects and from a slightly different point of
view: Lammeraner and Stafl [1.26], Stoll [1.39] and the most recent by
Krawczyk and Tegopoulos [1.25].
There is a vast literature on implementation of numerical methods to
solutions of magnetic field problems. It is not practical to offer a compre-
hensive bibliography within the compass of such a book as this one. A wide
range of papers is published in professional journals, such as the IEEE Trans-
actions on Magnetics, the lEE Proceedings A and B, the Journal of Applied
Physics, the Journal of Physics D (Applied Physics), and the International
Journal for Computation and Mathematics in Electrical and Electronic Engi-

neering, COMPEL. Relevant publications may also be found in the IEEE
Transactions on Power Apparatus and Systems, the International Journal
for Numerical Methods in Engineering, the Computer Methods in Applied
Mechanics and Engineering, and Computer Aided Design.
Much of the current material in any rapidly developing area is presented
at professional conferences. Those relevant to computational magnetics are
COMPUMAG, INTERMAG, MMM, CEFC, CEM, ISEF and ELECTRO-
SOFT. Selected and refereed papers from many of the above conferences often
appear in one of the professional journals, see for example proceedings of
ISEF'91 [1.42].
2
Finite-Difference Methods
Richard L. Stoll and Kazimierz Zakrzewski
2.1 INTRODUCTION
Finite-difference methods are important for two reasons. First, they form the
background to almost all later developments. Secondly, a finite-difference
method is relatively easy to construct and program to solve a particular
problem, or class of problems, that may not be suitable for an existing general
purpose software package using, say, finite elements, and so is ideal for the
relatively rapid development of special purpose computer programs.
The finite-difference approach to the numerical solution of field problems
is basically simple in concept, but considerably harder to examine theor-
etically in terms of accuracy and stability. To solve any field problem we must
first select the dependent variable (field component or components, poten-
tial etc.). This variable is a continuous function within the space and time
domain of the problem, and subject to appropriate boundary conditions (see
Chapter 1). The application of the finite-difference technique is then essen-
tially one of overlaying the problem domain with a net, grid or mesh that is
uniform in size but may alter from one sub-domain to another. The variable
now assumes a set of discrete values at the junctions or nodes of the mesh,
but remains undefined at intermediate points. There is, of course, a lack of
preciseness in this definition. For example, if the function !(x) is defined
by its values at the discrete points x = 0, tu, 2tu, ... , ntu, the function
h(x) =!(x) + g(x) sin (nx/ tu) is similarly defined but has different values
elsewhere. However, we aim to deal with smooth functions in the sense that
!i.x must be sufficiently small to allow !i.f! tu to be a good approximation to
d!/dx; something that would not be true for h(x).
The equations to be solved by the finite-difference techniques described here
are the second-order partial differential equations classified under their math-
ematical types: elliptic and parabolic, in Section 1.3. The actual elliptic equa-
tions of interest are Laplace, Poisson and Helmholtz which describe boundary
value problems. Parabolic equations on the other hand are, in our context,
the diffusion equations «1.80) and (1.81», which describe initial boundary
value or marching problems in the sense that the solution marches out from
the initial state, at time t = 0 say, guided and modified in transit by the sur-
rounding boundary conditions.
36 Finite-Difference Methods
2.2 DISCRETIZATION, FINITE-DIFFERENCE OPERATORS

AND TRUNCATION ERRORS
Consider a function f(x) which is a continuous function of x in the domain
(a, b). The domain X = b - a is now divided into M elements of equal length
lu, thus defining a set of M + 1 nodes at discrete values of x given by
xi=a+i~x, O~i~M (2.1)
The function f(x) is now represented by its values at these nodes,
/; = f{Xi) (2.2)
Between, nodes f(x) can only be found by interpolation. For example, using
linear interpolation in the ith element to find an approximation f' to f(x '),
where Xi~X' ~Xi+" we have
f' = &/; + 1 + (I - &)/; (2.3)
where
x' -x·
&=--' (2.4)
lu
The error in f' is given by the difference f(x ' ) - f' and is dependent on
the interpolation technique used. However, in practice, in the computation
schemes of interest here, the values of /; are not the exact analytically
determined values given by eqn (2.2) but are themselves found using a finite-
difference process. Thus a study of errors and their possible propagation is
important. To this end it is sometimes helpful to represent f(x) or its error
in terms of a Fourier series in complex exponential form
b
00
f{x) = gnej27tnxlx (2.5)

n= -00
where j = .J- 1, and
gn = ~Jf{X) e-j27tnxIX dx (2.6)
Equations (2.5) and (2.6) can be expressed in discrete form as
/; = b
M
Gnej27tnilM (2.7)
n=1
(2.8)
Once discretization has taken place the number of harmonics are clearly
limited. The wavelength of the smallest harmonic is equal to lu.
We not only need to discretize functions but also to derive suitable finite-
difference approximations to the first and second derivatives appearing in the
A One-Dimensional Parabolic Finite-Difference Equation 37
field equations. The Taylor series expansion is one method that can be used,
whereby, writing h = fix for brevity
f{x + h) = f{x) + hj'{x) + 1h 2r{x) + i h 3fm{x) + . . . (2.9)
f{x - h) =f{x) - hf'{x) + ih 2r{x) - ih 3fm{x) + ... (2.10)
Subtraction of these expansions yields
f{x + h) - f{x - h) = 2hj'{x) + O{h 3 ) (2.11)
where O(h 3) indicates terms containing third and higher powers of the small
length h. Thus
(2.12)
If O(h 2) is omitted from eqn (2.12), we say that the remaining difference
formula for f'(x) has a truncation error of order h 2• A difference formula,
with similar error, can be obtained for the second derivativer(x) by adding
eqns (2.9) and (2.10):
r{x) =f{x + h) - 2f~~) + f{x - h) + O{h 2 ) (2.13)
It is sometimes necessary, e.g. when dealing with boundary conditions, to use

eqn (2.9) to obtain a first derivative of the form
f'{x) =f(x + h} - f(x} + O{h) (2.14)

h
This is the so-called forward-difference formula (a backward-difference
equivalent comes from eqn (2.10», but the penalty is the first-order error
involved. However, formulas of the type given in eqn (2.14) are frequently
used for time-stepping diffusion equations.
2.3 A ONE-DIMENSIONAL PARABOLIC

FINITE-DIFFERENCE EQUATION
At this stage it is probably helpful to examine some possible finite-difference
formulations of a simple one-dimensional magnetic field problem that will be
solved in more detail later. One-dimensional problems are only significant
from the numerical point of view if either the excitation is of too complicated
a form to be easily handled analytically, or, the excitation is simple (e.g. varies
sinusoidally with time) but the response is nonlinear. We will consider the
latter as being the more general and take the example of a large plate of
ferromagnetic material immersed in a uniform time-varying magnetic field,
e.g. Hz = H s cos rot, applied parallel to the surface of the plate. Let the plane
of symmetry of the plate or lamination be on the plane y = 0 as shown in
Fig. 2.1. If the plate thickness 2b is much less than the other dimensions,
@ Hscoswt
~ ~ ~ ~ ~ ~ ~ ~ ~ ~
b
-- - - ~ - ~
b
~ ~ ~ ~ ~ -~-~ - .- -- - - ~ - ~ ~ -
z
~ ~ ~ ~ ~ ~ ~ ~ ~ -~ ~ ~ ~--_._.~ --_.- -- -- - - - - - - ~.~.~._~
x
@ Hs coswt
(a)
M
,t b H= Hscos wkt:J.)
y=, t
M-l
i+l,k
i,k-1 4 i,k -----1 i,k+l
i-l,k
3
2
1
0 /~ 1_ 2 3 4 5 6 k
-
y-O, aH/iJy-O
(b)
Fig. 2.1 (a) Ferromagnetic plate immersed in uniform magnetic field Hs cos wt, (b) half
thickness of plate modelled in the space-time domain
the problem can be modelled by a one-dimensional diffusion equation of the

form
a2H
-= 0--
dB aH (2.15)
ay 2 dB at
where H is an abbreviation for Hz(Y, t) inside the plate. Equation (2.15) can
be obtained from eqns (1.1), (1.2) and (1.8) by neglecting displacement current
(negligible in conducting materials) and assuming the conductivity 0 to be a
constant. Note that the constitutive equation (eqn (1.7» has not been used
because Band H are related by the magnetization curve, or a hysteresis loop,
as illustrated in Fig. 1.1. For brevity, however, let us define a parameter ~
as
where ~d is the differential permeability and a function of both y and t. Thus

eqn (2.15) becomes
o2H = poH
(2.17)
oy2 at
By symmetry the condition at the centre of the plate, y = 0, is that oH/oy = 0,
and only half the plate (0 ~y ~ b) need be considered and divided into M
mesh elements, say, where M = bl ~y. H is thus defined at M + 1 nodes in
space at the points y = i~y, where i = 0, 1, 2, ... , M and this space 'ladder' is
marched forward in time t taking discrete steps At so that t = kAt, where
k = 0, 1, 2, . . .. A given node in this discretized space-time domain can be
addressed using the integers i and k, so that H at node (i,k) is denoted by
Hi,k'
From the brief introduction to finite-difference formulae for first- and
second-order derivatives at the end of Section 2.2, we might suggest the
following centre difference algorithm to represent eqn (2.17)
Hi+l,k - 2Hi,k + Hi-l,k _ pHi,k+l - Hi,k-l
(2.18)
~y2 - 2At
However, this deceptively logical equation, first proposed by Richardson
[2.41] before the days of high-speed computers, is unconditionally unstable
and serves as a warning that the matter of convergence and stability cannot
be taken lightly. These topics are briefly considered in the next Section.
The next obvious alternative is a simple explicit two-time step algorithm
based on the forward-difference formula in time (see eqn (2.14». Thus
Hi+l,k - 2Hi,k + Hi-l,k = pHi,k+l - Hi,k (2.19)
~y2 At
giving
Hi,k+l = rHi+l,k + (1 - 2r)Hi,k + rHi-l,k (2.20)
where
At
r=-- (2.21)
p~y2
It can be shown (see Section 2.4) that r ~ 0.5 to ensure convergence and
stability in the linear case. In practice this means using very small values of
time step. In complete contrast, the backward-difference formulation in time,
yields
-rHi+I,k+l + (1 + 2r)Hi,k+l - rHi-I,k+l = Hi,k (2.22)
which is stable for all values of constant r, Le. in the linear case. The penalty
to be paid for the likelihood, even in a nonlinear system, of good stability,
is that eqn (2.22) must be solved simultaneously with all M equations that
apply to time step k + 1. The coefficient matrix is tri-diagonal and a simple
form of Gaussian elimination can be used [2.41J given in a very neat form
by the Thomas algorithm [2.51J.
The truncation error of both eqns (2.20) and (2.22) is O(~y2 + M), Le.
they suffer from a first-order error in M as we might have anticipated from
eqn (2.14). In the search for a smaller error consider a weighted average of
the second-order space derivatives on time rows k and k + 1, so that
r{0(Hi+I,k+1 - 2Hi,k+1 + Hi-I,k+.) + (1 - 0) (Hi+I,k - 2Hi,k + H i- I•k ))
= Hi,k+ J - Hi,k (2.23)
Clearly if 0 =0 we have the simple explicit equation, and if 0 = 1, the simple
implicit equation. Equation (2.23) is unconditionally convergent and stable for
Yz ~ 0 ~ 1, but for 0 ~ 0 < Y2, we have the condition [2.41J
r~ 2(1 - 20 ) (2.24)
However, for constant r, the minimum local truncation error occurs if [2.2J
0=!(I-~J
2 6r
(2.25)
Equation (2.25) suggests that, for fairly large values of r or values that are
variable, the sensible choice of 0 is 1/2, Le. on the boundary of unconditional
stability. Equation (2.23) then reduces to the Crank-Nicolson form
-rHi+I,k+1 + (2 + 2r)Hi,k+1 - rHi_I,k+1 = rHi+I,k + (2 - 2r)Hi,k + rHi-I,k
(2.26)
with error O(~y2 + M 2). This algorithm is a popular one because of its
stability and higher accuracy, which effectively means the ability to use a
larger time step for the same accuracy.
A relatively large time step in a nonlinear problem with r a function of H
requires at least one step of an iteration process, and possibly more, to modify
ri before moving forward by M. Bearing in mind that the Crank-Nicolson
algorithm for Hi,k+ I (eqn (2.26» is symmetrically centred about the inter-
mediate node (i, k + Y2), the value of Hi,k+ 'I, is first found by a half-step
Crank-Nicolson process using the existing value ~ of the nonlinear parameter
at each node, so that
M
ri,k - --
- 2~~y2
(2.27)
Once Hi,k+V, is known, the new parameter to be used in the main step
forward is
~' = ~(Hi,k+I/2) (2.28)
and
M
ri, k = ~' ~y2 (2.29)
Implicit equations have the disadvantage of needing simultaneous solution of

all M unknowns at each time step, and each half time step in the Crank-
Nicolson scheme with nonlinear coefficients. Explicit algorithms, if conver-
gent and stable for relatively large values of r, would be an obvious advantage.
One well-known explicit method is that of DuFort and Frankel [2.15) which
simply replaces Hi,k in the basic explicit scheme of eqn (2.20) by the average
of the i, k - 1 and i, k + 1 values, i.e.
H _ Hi,k+1 + Hi,k-I (2.30)
i,k - 2
The DuFort-Frankel algorithm is thus a 3-step one of the form

2r
H i.k+1 = Hi,k-I + 2r + 1 (Hi+l,k - 2Hi,k-1 + Hi-I,k) (2.31)
where r = ri,k is still the value of the parameter at node i, k. The local
truncation error is O(~y2 + M 2 + M 2I ~y2) where MI ~y must tend to zero
as ~y and ~t separately tend to zero. In practice this means that
M <C ~~ (2.32)
~y
otherwise the term (MI ~y)2 a 2Hlat 2 is effectively being added to the right-
hand side of the original diffusion equation (eqn (2.17» making it an hyper-
bolic equation.
A brief investigation will show that the Crank-Nicolson scheme can· be
derived from the application of the simple explicit and implicit eqns (2.20)
and (2.22) on alternative time rows and then shrinking this double time-step
process to one step. In a similar way, the DuFort-Frankel method is equivalent
to the application of eqns (2.20) and (2.22) to alternate nodes. Thus if the
explicit equation is first used to step forward from alternate nodes on row k
to row k + 1, the application of the implicit equation to the remaining nodes
on row k + 1 is an explicit process because at node (i, k + 1), for example,
the flanking values at nodes (i - 1, k + 1) and (i + 1, k + I) have already
been computed. Not only this, but immediate substitution of Hi,k+1 into the
explicit equation for the next step forward to time k + 2 yields
H i,k+2 = 2Hi,k+1 - H i. k (2.33)
and, after treatment of all nodes in this way, we are ready to use the implicit
equation again on row k + 2. Thus, provided the intermediate nodal values
are not required, there need only be the cyclic application of eqn (2.22) (in
explicit form) and eqn (2.33), and the computation involved is reduced by
about 30070. It is clear that, regarding the space-time domain as a chess board,
only nodes of one colour need be recorded and this led to the name Hopscotch
for this algorithm when originally proposed by Gourlay [2.20).
Unfortunately, some circumstances may make it necessary to return to the

original DuFort-Frankel form of the algorithm but the separation of the nodes
into two interlaced sets, black and white, is an important concept because it
is a warning not to mix the sets when calculating the coefficients ~. Because
~ at the black node (i,k), for example, is used to obtain H at the white node
(i, k + 1), it must be a function of the average of Hi+ I,k and Hi-I,k' not of
Hi,k'
It is worth commenting on this, at first, surprising fact that two methods
(Crank-Nicolson and DuFort-Frankel) so closely based on the dual applica-
tion of simple algorithms with truncation error O( ay 2 + M) should yield an
increased second-order accuracy in time (M 2 ). The reason is that the simple
explicit method approaches the true solution from below and the simple
implicit method from above [2.38], i.e. the fundamental components of the
error are opposite in sign.
2.4 CONVERGENCE, COMPATIBILITY AND STABILITY

Sufficient reference has been made to the local truncation error which is
defined as the difference between the partial differential equation and its
approximating difference equation at any given point. The discretization
error on the other hand is the difference between the exact solution u of the
differential equation and the exact solution of the difference equation U. The
scheme is said to converge if U tends to u at all nodes as the space and time
steps tend to zero, sometimes subject to a link between the latter (e.g. as in
the case of the DuFort-Frankel algorithm).
Because convergence is difficult to investigate, the simpler concept of
compatibility (or consistency) is often used. The difference equation is usually
said to be compatible with the differential equation if the limiting value of
the local truncation error is zero as the space and time steps tend to zero. A
very important theory propounded by Lax [2.59] states that, given a properly
posed linear initial-value problem and a linear finite-difference approxima-
tion to it satisfying the compatibility condition, stability is the necessary and
sufficient condition for convergence. Thus it may not be necessary to conduct
the harder examination of convergence directly.
Stability has to do with the growth, or strictly the lack of growth, of errors.
All numerical processes generate errors, e.g. rounding errors, and the actual
computed solution of the difference equation N differs from the exact solution
U, so that the total error at node (i, k) say is
Ui,k - Ni,k = (Ui,k - Ui,k) + (Ui,k - Ni,k) (2.34)

where the first term on the right-hand side is the discretization error and the
second term is the global rounding error R. In contrast to the discretization
error, the rounding error does not decrease as the mesh lengths decrease,
rather it will get worse because more arithmetic operations are involved.
Digital computers hold numbers to many significant figures and so rounding
errors are basically small. Nevertheless, for fixed values of the small mesh
Convergence, Compatibility and Stability 43
lengths, it is possible for the time-stepping process to allow the errors to

swamp the true solution.
One technique for investigating the behaviour of errors is the Fourier
stability method where a space row (or space grid if the problem is two-
dimensional) of errors is inserted in the form of a Fourier series with as many
terms as there are nodes in the row. In a linear system it is sufficient to
consider one harmonic only (because superposition can be used) and con-
sider its propagation as time t increases. Consider the nth harmonic, of unit
magnitude for convenience, so that the error E at t = 0 at node (i,0) is, from
eqn(2.7),
E i,O -- ei21tnilM -
-
eiAih (2.35)
where A = 21tn/Mh and h = !u for brevity. The errors that are now carried
forward in time are subject to the same difference equation as the exact
solution U. If we suppose that the method of separation of variables can be
applied to the difference equation, it is possible to write E at any node (i, k)
as
(2.36)
where p = Ii.t for brevity and y is a constant. The error will not increase with
time, Le. k, provided
(2.37)
As an example, let us submit these error functions to the simple explicit
equation. Substituting from eqn (2.36) into eqn (2.20) and dividing by Ei,k we
obtain
e YP = re iAh + (1 - 2r) + re- p·h
= 2r cos Ah + (1 - 2r)
=1- 2r (1 - cos Ah) =1- 4r sin 2 (A;J (2.38)
Since r is positive, eYP is always less than unity. There are three conditions
possible for the error solution (eqn 2.36):
(l) if 0 ~ e YP < 1, the error decays steadily as k -+ 00;

(2) if - I < e YP < 0, the error decays in magnitude but oscillates in sign;
(3) if e YP < - 1, the error oscillates with increasing magnitude, and the
solution is unstable.
The explicit equation is thus stable if e YP ~ - 1, or 1 - 4r sin 2 (Ahl2) ~ - 1 so

that
r ~ [2 sin 2 (Ahl2)]-1 (2.39)
The sine term cannot exceed unity, at which point r ~ Y2 as stated in Section
2.3.
Consider now the DuFort-Frankel algorithm. Writing eqn (2.31) in terms

of the error functions (eqn (2.36», dividing by E; k and putting the factor
2r(2r + 1) -I as a for brevity, we obtain .
e YP - e- YP = a (e jlJr - 2e- yh + e- jlJr ) (2.40)
Let K = e YP so that eqn (2.40) reduces to the quadratic
K2 - 2aX cos ')..h + (2a - 1) = 0 (2.41)
with the solution
K = acos')..h ± [a 2 cos 2 ')..h - (2a _1)]1/2 (2.42)
If, as stated in Section 2.3, the algorithm is unconditionally stable, r can take
any positive value. When r is large a tends to unity from below, and in the
limit
K = cos /..h ± j sin /..h (2.43)
Thus K ~ 1 as required (see eqn (2.37».
The Fourier stability method is the least rigorous of several techniques of
stability analysis, not least because boundary conditions are ignored. For a
more detailed survey the reader is referred to excellent books by Smith [2.41]
and Ames [2.2]. In particular, the stability and convergence of the DuFort-
Frankel algorithm has been investigated in detail for some one- and two-
dimensional nonlinear magnetic field problems by Wiak [2.53-2.56].
2.5 CONVERGENCE OF NONLINEAR

DIFFUSION EQUATIONS
The detailed analysis of Wiak on the DuFort-Frankel scheme mentioned at
the end of Section 2.4 is beyond the scope of this book, but it is worth
attempting two things briefly here. First to list various forms of the one-
dimensional diffusion equation in terms of three selected dependent vari-
ables Hz, Ex and Ax (using the same model configuration as in Section 2.3)
with the magnetic or electric material parameter as a variable (but excluding
dielectric materials). Secondly, to examine the convergence of a nonlinear
version of the basic explicit algorithm (eqn (2.20» as the simplest example of
the difficulties to be faced if the question of convergence is to be addressed
theoretically.
From eqns (1.1), (1.2), (1.7), (1.8) and (1.43) we have the one-dimensional
first-order equations,
(2.44)
(2.45)
Convergence of Nonlinear Diffusion Equations 45
aA x
ay = - Bz = - ~ ( H ) Hz (2.46)
The permeability expressed as some suitable function of H can represent either

a single-valued magnetisation curve or an hysteresis loop. The conductivity
o expressed as a function of E models one form of nonlinear conductor.
Other possibilities include variation with temperature, which would require
the coupling of the thermal and electromagnetic systems in some form of
interactive hybrid scheme, or the addition of H to give o(E,H) as proposed
in a recent model of a high-temperature ceramic superconductor [2.34].
Let us now assume 0 to be a constant and differentiate eqns (2.44) to (2.46)
in turn with respect to y and isolate the corresponding dependent variable
Hz, Ex and Ax by suitable combination and manipulation, to yield
2
a Hz _ (H) aHz (2.47)
ay 2 - O~d at
2
a Ex _ (H) iJEx _ 0 _ a~AH) E aEx (2.48)
iJy 2 - O~d at + ~AH) iJHz x ay
a [ 1 iJAx] iJA x (2.49)

iJy ~(H) iJy = -oEx = Oat
since, from eqns (1.2) and (1.43), Ex = -aAx/iJt. Equation (2.47) is iden-
tical to eqn (2.15) already introduced in the simple model problem of Section
2.3.
Alternatively with ~ = ~d =constant, and 0 a variable:
iJ [ 1 iJHz] aHz (2.50)

ay o(E) iJy = ~at
2
iJ Ex _ ( ) aEx E iJo(E) (2.51)
iJy 2 - 0 E ~ at + ~ x at
2
iJ Ax
--=oE~-
() aAx (2.52)
iJy 2 at
The mathematical similarity between eqns (2.47) and (2.52), and eqns (2.49)
and (2.50) is clearly seen. The type of boundary condition will be an additional
consideration and may suggest one particular dependent variable, but, apart
from this, there is an obvious advantage in using equations of the simplest
form (eqns (2.47) and (2.52».
Consider now the general form of eqns (2.47)
u/ = eI>(u, Uyy) (2.53)
where the suffIx denotes differentiation with respect to tor y, and we assume
del>
-;-~c>O (2.54)
lJUyy
(2.55)
What follows is a simplified version of a proof of convergence based on the

maximum principle analysis and given in Ames [2.2]. The simple explicit
algorithm for eqn (2.53) is
Ui,k+1 -- Ui,k + ~t (fJ;Ui'k]

Ui,h ~y2 (2.56)
where
fJ;Ui,k = Ui+I,k - 2Ui,k + Ui-I,k (2.57)
On the other hand, from eqn (2.53)
Ui,k+1 = Ui,k + M Utli,k + 0(M 2)
= ui k
,
+ M (Ui k' 0;U i2,k]
, ~y
+ 0(M 2 + M~y2) (2.58)
Subtracting eqn (2.56) from eqn (2.58) the error function is given by
Ei,k+1 = Ei,k + M~ + 0(M 2 + M~y2) (2.59)
where
~ = fJ £. +
fJu I,k
fJ
fJu yy
[0; Ei , kJ
~y2
(2.60)
Thus, using eqn (2.57)
Ei,k+1
~t fJ
= uy
A
( fJ
2 -;;--Ei+I,k + 1 + ~t~ -
2~t fJ
A 2 -;;--
J Ei,k
uUyy uU uy uUyy
+-M2 -;;--E
fJ (2 2)
.:ly uU i- I, k + 0 M + M~y (2.61)
yy
All the coefficients of eqn (2.61) are positive if

fJ 2M fJ
1+ M---2-~0 (2.62)
fJu ~y fJuyy
or, using eqn (2.55), if
(2.63)
giving the condition,
M
~y2 ~
~I-Md=!(!-MJ=~
2d 2 d 2d
(2.64)
Subject to the condition expressed in eqn (2.64), eqn (2.61) can be written
in terms of the absolute values of the error function with lEi, k+ 1 I less than
Convergence of Nonlinear Diffusion Equations 47
or equal to the right-hand side. If IIEkl1 is the maximum of the absolute

values of Ei,k on time row k, then
IIEk+dl ~ (I + M~:J IIEkl1 +MA(M + Ay 2)

~ (1 + Md) IIEkll + MA (2.65)
where A(M + Ay 2) is some small quantity of order M and Ay 2. But IIEol1 = 0

and so, from eqn (2.65)
IIEIII~MA
IIE2 11 ~ (1 + Md)MA + MA
IIEk+dl ~ [1 + (1 + Md) + (1 + Md)2 + ... + (l + Md)k]MA
(1 + Md)k - 1 k A
= (I + Md) _ 1 MA = [( 1 + Md) - 1] d (2.66)
The error function thus tends to zero as M and Ay tend to zero. Given eqns
(2.54) and (2.55), this nonlinear form of the simple explicit equation is con-
vergent if condition (2.64) is satisfied.
It may be helpful to look at a specific case of nonlinearity. The modi-
fied Frolich representation of the magnetization curve is sometimes used,
whereby
H
B = 1l0H + a + blHI (2.67)
giving, from eqn (2.16),

dB oa
(2.68)
~ = 0 dB = Ollo + (0 + b1HI)2
Writing eqn (2.47) in the form of eqn (2.53)
aH = cI> = ~ a2H (2.69)

at ~ ay2
so that
(2.70)
and, with a = a + b IHI for brevity,
acl> a2H a
aH = ay2 aH
[IJ~ = aayH2
2 0(0
2aba
+ lloa2)2
aH 2ab
= at a(o + lloa 2) (2.71)
we have, from eqn (2.55),
d {I
~ (0 + ~oa2) aa~12:b + :2} (2.72)
If we assume that the maximum value of IaHlatl is roHs (as for sinusoidal
variation with time with peak value H s) and take typical values of the con-
stants in eqn (2.67) as a = 600 and b = 0.6, we can examine the condition
on the parameter d for different values of magnetic field. For very low H,
a =a, and with a good conductor (a > 106) the term a 21 a in eqn (2.72) is
=
negligible. Also a + Iloa 2 a, so that d ~ 2m IHI bla. With 50 Hz excitation
this means that d ~ 0.6281 HI. The maximum value of H to reasonably satisfy
the condition a = a is 50 A m -I, giving d min = 31.4, so that eqn (2.64)
becomes M ~ L\y 2/62.8 and M may have to be extremely small, depending
on the space discretization necessary. For values of H higher than about
50 A m -I, eqn (2.72) must be examined in more detail. However, for very
high magnetic field strength, a=blHI so that (a+lloa2)=(a+llob2IHI2)
and both terms in the factor may be important. As an example, suppose
IHI is such that llob2IHI2=a, then d~lIaH/atl/IHI + (2a llo)-I). One
problem is evident; namely that of finding a representative value of the ratio
IaH/at I/ IHI. Although both maximum values do not occur simultaneously
or everywhere in the plate, it is probably reasonable to take laHlatl/lHI =
roHslH s = ro, and, since the factor containing a now becomes negligible in
contrast, we have d ~ ro = lOOn. As a result, the modified Frolich curve does
not appear to be very promising when used with the simple explicit equation.
However, it must be remembered that the maximum analysis of convergence
is very pessimistic in nature. The form of the nonlinearity is also important.
For example, taking the curve represented by B = aHl/n, where n is an
integer, a wide range of H from low values upwards can be shown to require
d~ro(l-n-I).
2.6 FINITE-DIFFERENCE METHODS APPLIED TO

MAGNETIC FIELD PROBLEMS
There are numerous examples in the literature of the application of finite-
difference methods to magnetic field problems. The categories of problems
can be listed as follows:
(1) one- and two-dimensional (in space) solutions of the diffusion equa-
tion in various forms, with nonlinear coefficients and/or transient or
periodic nonsinusoidal excitation;
(2) two-dimensional solutions of the Helmholtz equation, where the
dependent variable is complex. By assumption this class is linear with
sinusoidal excitation;
(3) two- or three-dimensional solutions of Laplace's or Poisson's equations.
We will consider some of these possibilities in turn with examples.

Finite-Difference Methods Applied to Magnetic Field Problems 49
2.6.1 Steady-state eddy cu"ents in je"omagnetic plates, laminations

or cylinders
This type of problem was introduced in Section 2.3 where the applied magnetic
field is both uniform and parallel to the surfaces of a flat plate. With an
assumed constant permeability, the only solutions of interest involve transient
excitation or periodic excitation with some complex non-sinusoidal form.
Otherwise an analytical solution is straightforward.
Here we are primarily concerned with the nonlinearity introduced by the
magnetization curve of the material. Most examples in the literature con-
sider a simple sinusoidally varying field, H s sin rot, applied parallel to the
surface of a plate or parallel to the axis of a cylinder. Abrams and Gillott [2.1],
for example, consider the case of a long steel rod, and include the effect of
hysteresis.
The one-dimensional cylindrical form of the diffusion equation in Hz is
r
a2Hz 1 aHz aHz
ar2 + ar = Ol1d-at (2.73)
The simple explicit finite-difference form of eqn (2.73) is

H;+I .k - 2H;,k + H;_I 'k + . k - H;_I ,k = Ol1d'
H;+I H; k+1 - H;'
k.(274)
!i.r 2 2r!i.r !i.t
At the centre of the rod, there is a problem with the second term on the
left-hand side of eqns (2.73) and (2.74) as r tends to zero. However, from
Maclaurin's expansion
H'(r) = H'(O) + rH"(O) + ... (2.75)
and, therefore, since H'(O) = 0 (Le. aH/ar = 0 at r = 0), we have H'(r)/r
tending to H"(O) as r tends to zero. Thus eqn (2.73) becomes, for r = 0
2 a2Hz _ aHz (2.76)

ar 2 - °l1d at
and eqn (2.74) is
4(H;+I,k - H;.k) = 011 H;.k+1 - H;.k (2.77)
!i.r 2 d!i.t
It is interesting that Abrams and Gillott [2.1] appear to have succeeded in
using the simple explicit method. At about the same time Poritsky and Butler
[2.32], considering both thick flat slabs and thin plates, found some instabi-
lities and turned to the simple implicit scheme (see eqn (2.22», but with an
iterative solution at each time step based on an initial set of new values extra-
polated from the two previous time rows.
The most general case of periodic H boundary conditions at the surfaces
of a ferromagnetic plate is:
at y=O
q
Hz = H o + ~ HI sin (wit + </>J (2.78)
1=1
at y= 2b
q
Hz = Ho + ~ H/ sin (wit + </>/) (2.79)
1=1
where H o and H o are constant d.c. field components. At t = 0, 0 < y < 2b,
the initial value of Hz is usually set to zero.
The lack of symmetry indicated by the prime notation allows a component
of imposed transport current to be present, e.g. if H/ = - HI we have the
situation of imposed current only and no ambient magnetic field. In finite-
difference form the equivalent of eqn (2.78) for the surface y = 0 is
q
HO,k = Ho + ~ Hlsin (wlkM + </>/) (2.80)
/=1
and similarly for H 2M,k on the surface y = 2b, where M = bl~y. Only in
the two symmetrical situations can the finite-difference solution be con-
fined to half the plate. For odd symmetry (H/ = - HI) we have Hz =0, Le.
HM,k =0 at y =b, and for even symmetry (H/ =HI) aHz/ay =0, giving
HM+1,k=HM-1,k so that the equation for HM,k+1 can be rewritten with node
M + I eliminated. For example, the DuFort-Frankel equation (eqn (2.31)
becomes
2r
HM,k+1 = HM,k-1 + 2r + I (2HM_1,k - 2HM,k-l) (2.81)
The alternative to an imposed surface magnetisation is an imposed flux

per unit width of the plate. If the excitation winding producing the magnetic
flux in the plate or stack of laminations is connected to a sinusoidally vary-
ing emf source, Faraday's law indicates that the total flux in the plate and
the electric field on the surface of the plate are also sinusoidal. Since, by
symmetry,
a<l» a
= - at
- = --
A
2E (<I» sinwt) (2.82)

S at
where Es is the instantaneous value of the surface electric field, we have
Es = - ~ wei> cos wt (2.83)
But
aH
Jx = oEx = ayz (2.84)
and so we have the surface condition on y = 0 in finite-difference form

HI k - H I k 1
, 2~y - , = -2(O<JCIl cos «Ok~t)
A
(2.85)
The fictitious node (- 1, k) in the finite-difference equation for H o,k+ I can

therefore be eliminated. At the centre of the plate, node (M, k + 1), an
equation of the form of eqn (2.81) still holds because Ex = O.
The surface, and possibly the interior, values of the electric field are also
of interest when surface H is specified. Three-point polynomial fits, with error
O(~yz), for the surface nodes are given by [2.11]
E _ -3Ho,k + 4HI.k - H Z.k (2.86)

O,k - 2<J~Y
E - HZM-Z,k - 4HZM- I ,k + 3HZM.k (2.87)

ZM.k - 2<J~Y
In the symmetrical cases, with only half the plate computed, E ZM.k = ±EO• k
as appropriate.
The most important use of the surface electric field is in the calculation of
the total loss, both ohmic and hysteresis, per unit surface area of the plate.
From Poynting's theorem, the loss is obtained by integrating the product
(E x H) . n on both surfaces of the plate over a complete period T of the
time cycle, where n is the unit vector normal to the surface of the plate and
outwardly directed. Thus n = -jon the surface y = 0 and n = j on y = 2b,
giving
n
t+T
P =~ J ([iEAo) x kIiAO)] . (- j) + [iEA2b) x kIiA2b)] . dt

t
t+T
=~ J [EAO)Hz(O) - EA2b)HA2b)] dt (2.88)

t
Once the solution, in terms of the nodal values of H has converged, which
will take several cycles, one final cycle can be computed during which time
the surface nodal values of E can be obtained from eqns (2.86) and (2.87),
and the loss obtained by numerical integration using either the trapezoidal rule
or Simpson's rule. In terms of the latter, the first part of eqn (2.88) becomes,
with T=NM,
1
PI = 3N (Eo,nHo.n + 4Eo.n+IHo.n+l + 2Eo.n+zHo.n+z
+ 4Eo.n+3Ho.n+3 + ... + 4Eo.N+n-IHo.N+n-l + EO.N+nHo.N+n)
(2.89)
where n is the initial value of the time integer k for the final cycle.
The solution marches forward from the initial nodal values of H i • O = 0
at k = 0 and progresses through a numerical transient that may also be close
to the real transient if the initial conditions are realizable. However, the
important requirement is convergence to the steady state, which can be

checked at the end of each cycle by a simple test comparing Hi,k with Hi,k-N'
For example if
IHi,k - Hi,k-NI <& (2.90)

for all values of i, where & is some small positive quantity, the solution has
converged. & can be found by numerical experiment and will depend on the
results being sought.
1.6.1 Treatment of magnetic nonlinearity and hysteresis

The main, indeed the only, difficult problem with these one-dimensional
solutions concerns the variable coefficients ri, k that contain the differential
permeability !J.d' Ferromagnetic materials are subject to hysteresis (see
Section 1.5). Strictly speaking therefore, the hysteresis loops need to be
modelled in some way, although the single valued magnetization curve is
often used when modelling hard materials such as silicon steel laminations.
The additional problem is that, with cyclic magnetization, the operating loop
is a function of the peak value of Bor H. Clearly it is not possible to store
a large number of loops and so some means of interpolating between loops
is required.
There are two forms of major hysteresis loop. The first is the overall loop
formed by plotting the average flux density in a plate or lamination against
the surface value of magnetic field strength for a given fundamental fre-
quency. This is known as the dynamic loop because it contains the effect of
eddy currents, and can only be obtained once the full numerical solution is
known. The second type is the static hysteresis loop relating local values of
Band H and obtained by measuring successive small changes in B produced
by small and slow changes in H (see Section 1.5). Clearly, since domain wall
movements are involved, the peak to peak swings in H must be slow enough
not to produce significant additional microscopic eddy currents induced by
rapid domain wall movement. In the numerical solution it is necessary to
neglect the latter and assume that the static loops apply within the material.
As the result of normal macroscopic eddy-current screening the peak value of
H will decrease with depth from the surface and a different loop will apply
at each level.
Zakrzewski and Pietras [2.62] describe a method in which, starting with
the initial condition Hz = 0 at all nodes, the solution moves up the peak
magnetization curve from the demagnetized state. This curve is found experi-
mentally by joining the tips of a set of major hysteresis loops, and is almost
identical to the normal initial magnetization curve. Once the local maximum
value of Hi,k has been reached, the solution moves back down the upper
branch of the appropriate major hysteresis loop. When the maximum negative
value of Hi,k occurs, corresponding to some value B m of flux density, the
operating point is moved horizontally across to the reversal peak magnetiza-
tion curve (i.e. selecting the corresponding value Bm,Hm on the latter curve)
B (T)
0.8 ~_-7Bl
B2
-l-~~-rl""T--- B 3
-5
-B 1 -~~---1;/
-B 2 ---,~---
-B3
Fig. 2.2 Illustration of the upper and lower branches used in the computation process to
form a set of major hysteresis loops
in order to join the lower branch of another major hysteresis loop. The same
process repeats when the maximum positive value of H iok is obtained. After
a few cycles the value of H at each node will be found to be following the
symmetrical major loop appropriate to the peak field strength achieved at that
node. This numerical transient is similar to the actual transient process that
might occur in practice.
As already mentioned, only a few major hysteresis loop branches can be
stored as tables or files of Hp,Bp number pairs, where p represents the entry
number. However, the approximate profile of any intermediate branch can be
obtained by linearly interpolating according to the ratio
R = Bm - Bu (2.91)
Bu - 1 - Bu
where B u and B u - I are the peak values of flux density of the branches on
either side of B m • In order to find R the stored H,B entries in each branch
file must be artificially extended to the maximum value of H for the largest
hysteresis loop of the complete set, as illustrated in Fig. 2.2. Assuming that
the current value of Hi,k corresponds to an entry, say H p , in the files, the
required value of B is given by
B = Bu,p + R (Bu_I,p - Bu,p) (2.92)
where Bu,p is the pth flux density entry in the file for branch u. In practice,
Hi,k will not match a value H p exactly, but to obtain Ild( =!J.B/!J.H) we
actually require the values of Band H on either side of the point corresponding
to Hi,k'
Further details of the technique are given by Zakrzewski and Pietras [2.62]
who also show two examples; one for a thick slab in which hysteresis is
important at low excitation levels and one for high excitation in a thin lami-
nation where hysteresis loss may be comparatively important because of the
400 ,,-,,
,,
I
0-75 I ,
0'50
,,
200
\
\
0·25
N \
-t E
~o 0 -t
~ 0 1W7I'---'-------Il~~----JL---_W..,.,'--
\
I ~ I
wt I
I
I
I
-200 I
-0'50 - I
I
I
I
-0·75 I
\ I
-400
'-"
Fig. 2.3 Eddy-current density waveforms at different depths below the surface of a large
flat solid iron slab with a peak surface field of 410 A m -I: (a) 0 mm (b) 0·25 mm
(c) O' 50 mm (d) O' 75 mm
reduced eddy-current flow (and loss) due to saturation. Figure 2.3 illustrates
the eddy-current density as a function of depth from the surface of a slab of
solid iron on which the peak applied field is 410 A m -I. The frequency is
50 Hz and conductivity 5.4 X 106 S m -I. ~y = 0.125 mm and time steps in the
region of 80 IlS were used. The computed surface density agrees well with
measurement on the surface of a mild steel cylinder inside a solenoid. The
computed total loss per unit surface area is 20.2 W m -2, whereas without
hysteresis the loss is about 250/0 less.
Other attempts to represent hysteresis include semi-analytical models, e.g.
Coulson, Slater and Simpson [2.10] where flux density errors less than 10%
are claimed, and Del Vecchio [2.14], both based on the Preisach domain
model. On the other hand, Rivas [2.36] is typical of a good fully analytical
model using a rational fraction of the form of eqn (2.97) (see below) which
has also been employed for single-valued magnetization curves.
In fact it is more usual to neglect hysteresis and work with the peak
magnetization curve. The latter can be represented as a single table of B, H
points. Any value of B can be found by linear interpolation, i.e. if H lies
between the table entries H p and H p + I' we have
H-H
B = Bp +H _ ~ (Bp + 1 - Bp ) (2.93)
p+1 P
Clearly, it may be more convenient to list the pairs (Il, H), (Ild, H), (Ild, B),
(v,B 2 ), where v is the reluctivity, depending on the form of the finite-

difference equation to be solved. The alternative is an algebraic function
expressing B in terms of H. Fischer and Moser [2.17] have surveyed a large
number of possibilities, and Trutt, Erdelyi and Hopkins [2.49] examine some
in terms of the goodness of fit to actual measured curves. A simple equation
that is recommended for materials with a relatively smooth knee to the curve,
Le. transition into saturation, is the modified Frolich equation
H
B = JJ.oH + a + blHI (2.94)
which leads to
JJ. = JJ.o + (a + bIHI)-1 (2.95)
and
JJ.d = JJ.o + a(a + blHI )-2 (2.96)
A rather more complicated expression is the rational fraction approximation
of Widger [2.58]
2
B = ao + a,IHI + a21 H I H (2.97)
I + b,IHI + b21HI2
where higher order terms, Le. in H 3 etc., can be added. However, the
preferred form is that proposed by Rivas [2.36] where eqn (2.97) is taken to
represent the magnetization vector M not B (see eqn (1.98» so that
2
B- H ao + adHI +a21 H 1 H (2.98)
- JJ.o + I + bdHI + b21H I2
A short but useful discussion of the treatment of nonlinearity is also given
by Krawczyk and Tegopoulos [2.25]. The clear advantage of eqns (2.94) to
(2.98) is the limited computation involved. Lim and Hammond [2.28; 2.29],
for example, have used the original Frolich formulation (without the term
JJ.oH in eqn (2.94», together with the DuFort-Frankel algorithm, for exten-
sive studies of the effect of saturation on the eddy-current losses in steel
plates.
A much more complex expression which gives an excellent fit for silicon
steel laminations with a sharp knee in the B/H curve is
B = JJ.oH + 8, tan-' (A,H) + 82 tan-I (A 2H) (2.99)
where, as a guide, typical values of the constants are 8 1 = 0.987, 82 = 0.323,

A, = 0.030 8, and A 2 = 0.000 110.
2.6.3 Transient one-dimensional solutions

Transient solutions of the one-dimensional plate problem are considered by
Wiak [2.53; 2.54] who, as already mentioned, pays detailed attention to the
DuFort-Frankel algorithm. In reference [2.53] he formulates the problem
in terms of B, having chosen to specify the surface value of B in the form

Bm {1-exp(-th)} where B m is the final constant value and 't the time
constant. A small value 't = I ms gives something close to a step-function
impact excitation, and, especially with a small space step of 0.1 mm for a fairly
thin 5 mm plate, requires a very small time step of 1 IJ.S to follow the transient
solution accurately. In fact, Wiak specifies the inequalities
(~;J~ 0.01 and [~J ~ WOO (2.100)
Not surprisingly, the requirements for a transient solution are more severe
than with periodic steady-state excitation where the initial transient is not of
interest.
It is important to model the start of a real transient correctly; something
that is not possible using a three time-step scheme. Wiak proposes the two-step
Saulev algorithm [2.53] which has an asymmetric form with the second deriva-
tive approximated by (for the first step from k =0 to k = 1)
B·I, 1 + B'-II
2
a B B'+ " l
0 - B·I, 0 - "
(2.101)
ay 2 "" ~y2
The Saulev scheme is normally used with opposite skew on alternate time
rows, but here, from k = 2 onwards, the DuFort-Frankel equation takes
over.
It is not clear why Wiak uses the B formulation, which, although the basic
one-dimensional diffusion equation is not listed in Section 2.5, can easily be
derived and shown to be more complicated than the equivalent eqn (2.47)
for H. After all, in one class of one-dimensional diffusion problem (current
driven), surface H is the natural boundary condition. However, in reference
[2.54], Wiak examines the same transient solution in the E formulation (see
eqn (2.48». His conclusions about M and Ax are the same as in eqn (2.100).
For the situation where a plate is immersed in a magnetic field provided by
a voltage-driven excitation winding, the specification of E follows from the
fact that the surface electric field is related to dCl>/dt, so that we have the
transient equivalent of eqn (2.82). This fact does not necessarily make the
more complicated E equation preferable to the H equation, but it is likely to
be so because of the Dirichlet boundary conditions. As already noted in
Section 2.6.1, the H formulation here requires the surface Neumann condition
aH/ay (= aEx )' Alternatively, the vector potential formulation (eqn (2.49»
can be used because, from curl A = B,
b b
CI>{t) = fBzdY = -fa;xdY =As{t) (2.102)

o 0 Y
where As is the instantaneous surface value of Ax, and Ax at the centre Y = b
is assumed to be zero.
Stoll [2.43] gives another example of transient current density patterns at
three levels inside a 5 mm steel plate when the surface H field is suddenly
A=O ~
z
Fig. 2.4 Low and high voltage coil cross-sections in half the window of a single-phase
transformer
switched on or off. There is an interesting difference in the magnitude of the

current response under the two regimes, not due to hysteresis because this is
not included, but to magnetic saturation. The current density is always greater
at or near the beginning of the transient, but, due to saturation, the rate of
fall of flux is less than the unsaturated rate of rise.
2.6.4 Two-dimensional static solutions in multi-region problems

Two-dimensional problems involving Laplace's or Poisson's equations are
often amenable to analytical solution (see, for example, the excellent survey
by Binns, Lawrenson, and Trowbridge [2.4]) unless the boundary is of com-
plicated shape or there are two or more regions involved, particularly if the
latter involve magnetic materials.
The most complicated Poissonian type of multi-region problem solved
analytically uses a method devised by Roth, and one transformer example is
outlined by Binns et al. [2.3; 2.4]. The disadvantage of the method is that the
field quantities, and resulting parameters such as the force on a conductor,
appear as double Fourier series. Thus the computation involved is substantial.
In addition, the conductor cross-sections must be rectangular and the current
density uniform. The alternative of a finite-difference or finite-element solu-
tion soon becomes attractive.
A typical configuration is given in Fig. 2.4 which shows half the window
of a transformer and the cross-section of the low and high voltage coils. The
plane of symmetry is shown as a dotted line. The two-dimensional Poissonian
equation to be solved is
a2A a2A
ax/ + ay 2 = -llo Jz (x,y)
Z
(2.103)
M
1!.1Y
~
i-l,j
i,j+l
i,j i+l,j
iJ-l
}--,
z
N
Fig. 2.5 Finite-difference mesh representing the problem in Fig. 2.4
the vector potential being an obvious choice of dependent variable because A

and J are parallel. From eqn (1.43)
Bx = o~z and By = _ o~z (2.104)
and the boundary conditions in Fig. 2.4 follow because the transformer core
is assumed to be infinitely permeable so that there is no tangential component
of magnetic field at the window surfaces. On the plane of symmetry, Bx = 0,
so that oA/oy = 0 or A = constant. Only the derivatives of A are important
and the constant can be taken as zero for convenience.
The finite-difference grid representing Fig. 2.4 is shown in Fig.2.5. A
rectangular mesh is used with a uniform cell of dimensions t::..x x t::..y. Referring
to eqn (2.13), a finite-difference equivalent of eqn (2.103) can be written
immediately as
A i + I ,} - 2A i,} + Ai-I,} Ai,}+1 - 2A i,} + Ai,}-l _ J
(2.105)
t::..x2 + t::..y 2 - - Ilo i,}
where Ji ,} is the current density associated with node i, j. The nodal current
is assumed to flow with uniform density Ji ,} in a cell of dimension t::..x x t::..y
having node i,j as its centre. It is therefore a simple matter to write differ-
ence equations for nodes on the flat surface of a conductor, at a corner, or
even at an inverted (concave) corner such as occurs with an L-shape, by using
the weighted averages 0.5J, 0.25J, or 0.75J, respectively in place of J in
eqn (2.105).
With the Neumann boundaries, where oA/on is specified, the boundary
can either bisect the outer cells as illustrated in Fig. 2.5, or coincide with the
outer nodes. Treatment of the latter case is as given in the one-dimensional
problem of Section 2.6.1 where the fictitious outside node is eliminated from
the equation for the boundary nodes. In Fig. 2.5, a node outside the boundary
has the same potential as the corresponding node just inside, so that, for
example, in eqn (2.105) written for a node adjacent to the left-hand boundary,
we replace Ai_I,j by Aij, since aA/ax=O.
Apart from the nodes adjacent to the boundaries, all internal equations have
the form, from eqn (2.105),
~y2 ~X2
Ai,j = 2(~2 + ~y2) (Ai+1,j + Ai_I) + 2(~2 + ~y2) (Ai,j+l + Ai,j-l)
~2~y2
+ 2(~2 + ~y2) J!OJi,j (2.106)
It is clear that if the complete set of equations for all the unknown nodal
values was written in matrix form, the coefficient matrix would be extremely
sparse, having at most five non-zero entries per row, whereas the dimension
of the matrix even in the simple illustration of Fig. 2.5 is 195. The orderly
addressing system in Fig. 2.5 also means that the matrix entries form a com-
pact band surrounding the main diagonal. Normal matrix elimination methods
(e.g. Gaussian elimination) are not efficient in this case and it is preferable
to use an iterative method, or, for a simple uniformly meshed rectangular
region such as Fig. 2.5 with boundary conditions that are uniform on each
boundary, one of the so-called direct methods. These can be expected to give
at least one order of magnitude improvement in solution time over a good
iterative method.
Interestingly, .these direct methods closely resemble the original Roth
double-Fourier series analysis of a continuous variable but applied now to
a set of nodal potential and current density values, where the number of
harmonics is limited to the number of nodes in each direction. In 1965,
Hockney [2.21] described the application of Fourier expansions in one or
other, or both, coordinate directions, but advocated series expansions in one
direction only due to the time taken to determine the Fourier coefficients.
By substituting the series, in x say, into the original Poisson's equation, an
ordinary differential equation (in y) is obtained for each harmonic coefficient.
The resulting set of equations can be written in tri-diagonal matrix form and
solved numerically. A process known as cyclic reduction lessens the computa-
tion considerably [2.21]. Since the work by Hockney, several developments
have been reported [2.6; 2.26; 2.39; 2.40; 2.48], but still restricted to basically
rectangular and uniformly bounded problems.
A further method, applicable to both Poisson's and Helmholtz equation
in less restrictive geometries with graded meshes, is the Capacitance Matrix
technique (CMT) described by Proskurowski and Widlund [2.33]. A com-
bination of CMT and the earlier methods has been used by Lukaniszyn [2.30]
to produce the magnetic leakage field surrounding the windings of a current
transformer. Figure 2.6 shows the contours of constant vector potential
obtained for two two-dimensional solutions, the first (a) on the centre plane
of the core and window, and the second (b) perpendicular to the first and on
10
(a)
Fig. 2.6 Contours of constant vector potential in and around two transformer coils: (a) on
the centre plane of the window, and (b) perpendicular to the window and on the centre
plane of the core leg
the centre plane of a vertical core leg. Although to treat a three-dimensional

problem as a pair of orthogonal two-dimensional solutions in this way is to
take considerable liberties with the geometry, it is possible to obtain the leak-
age reactance within a few percent of the measured value.
Considering just one of the two-dimensional fields, e.g. Fig. 2.6a, the self
inductance per unit length of one coil with N uniformly wound turns is
(2.107)
where I is the coil current, 8 the cross-section, and A and A are the values
I
of vector potential at corresponding points in the two sections of the coil,

when only that coil is excited. The mutual inductance between two coils is
given by
(2.108)
where I. is the current in one coil, Al is the vector potential produced by that
coil, and the integration is carried out over the cross-section 8 2 of the other
coil. If L I and L 2 are the self inductances of the primary and secondary coils
respectively, then the leakage inductance referred to the primary side is usually
defined as [L I -M(L/L 2)V,).
In a regular mesh, where the conductor or coil side boundaries coincide with
the grid lines, the numerical integration is particularly simple, so that
H cIs "" ~l
A (Ai,j + Ai+I,j + Ai+1,j+1 + Ai,j+l) !U!i.y (2.109)
where the summation takes place over all the cells in the cross-section of one
coil side.
The force on a conductor or coil can be obtained in a similar way. In general
F = IH (J x B) dv (2.110)
Now in x,y coordinates,
z~ ':' oAz~ oA ':'

B=Bl+BJ=-I--J
x y oy ox
(2.111)
so that the two components of force on one of the coils in Fig. 2.6a are
(2.112)
Fy = II Jz oA
oy
Zdxd
y (2.113)
In the example considered here Jz is a constant and so we are left with the
need to integrate oA/ox etc. over the coil cross-section
II
oA 1
ox dxdy "" ~2 [(Ai+l,j - Ai) + (Ai+l,j+1 - Ai,j+l)]!i.y (2.114)
The alternative to direct methods of solving the sets of nodal equations

of the type found in eqns (2.105) and (2.106), and essential in more compli-
cated multi-region problems, are iterative methods of which successive over-
relaxation (SOR) is the most common. The potentials of all the unspecified
nodes are initially set to zero and we begin to scan the grid repetitively in an
orderly and systematic manner, forcing eqn (2.105) to be satisfied at each
node. One possibility is to scan the nodes of Fig. 2.5 by columns bottom to
top and from left to right, so that, during the (k + l)th scan or iteration, the
form of eqn(2.106) becomes
A (k+l)
i,j -- C 1A(k)
i+I,j + C2 A(k)
i,j+l +
C3 A(k+l)
i-l,j +
C4 A(k+l)
i,j-l +
C5 Ji,j (2 . 115)
where C l , Cz etc. are the known coefficients in terms of !U and !i.y etc. We
note that the new values of the potential are used as soon as they are available
and so only one array of values need be stored at anyone time. Equation
(2.115) expresses the Gauss-Seidel iterative algorithm. The extension to SOR
is simply to obtain the change that is about to be made to Ai,j (known as the
displacement or residual) Le.
R~k+l)
I,}
= A~k+l) _ A(k)
I,J I,}
(2.116)
and amplify it by multiplying by a relaxation factor a that is greater than
unity, to give
SCANS
, N-552
300 ,,
,,
,,
,,
11-132 ,,
200 ,
"- ,
,,
,,
, , I
, I
100
oL-.--L_--l-_.L...---l_--L_...I-_I.-...-l._--l-_.l-_
1.1 1.2 1.3 1.4 I.S 1.8 1.7 1.8 1.9 2.0
Fig. 2.7 The number of SOR iterations needed for a given minimum residual as a function
of the number of internal nodes (N) in a simple square mesh and the value of the relaxation
factor a
A(k+l) = A(k) + aR(k+l) (2.117)

',J ',J ',J
If a= 1, eqns (2.116) and (2.117) reduce to eqn(2.115) again. The SOR

iterations will converge if 1 ~ a < 2, but the solution is very significantly
speeded up in terms of the number of scans required, for a given error, if
a is close to its optimum value a o for the size and type of problem being
solved. Typically the number of scan will be an order of magnitude less than
that needed by the original Gauss-Seidel scheme.
For a uniform rectangular region 0 x b, Garabedian [2.19] has shown that
2K
a
o
= K + ~y.JK [(1t/O)2 + (1t/b)2] 1/2
(2.118)
where K = 1 + (~y/ lui. However, in more complicated problems, the

method of Carre [2.9], that deduces successive improvements to a during the
iterative process, approaching a o from below, can be very effective. Figure
2.7 illustrates the way in which the number of scans required decreases rapidly
near the optimum value and also the penalty of exceeding a o by even a small
amount.
The work of this Section has been limited to the numerical solution of
Laplace's and Poisson's equations, with uniform meshing, in non-magnetic
current carrying regions. Once we have mixed magnetic and non-magnetic
regions special finite-difference equations will apply at nodes on the inter-
faces. Although techniques exist to formulate these equations by applying the
magnetic boundary conditions (see, for example, Binns and Lawrenson [2.3]),
a more general approach is outlined in the following Section.
y 2
d ,.··············c.···················;b
i N
3 z 1
x
eJ W O·S E Ja
f ; 8" ·················'h
Fig. 2.8 A node in an irregular grid with the surrounding cells
2.6.5 More two-dimensional static and steady-state solutions

In developing a finite-difference equation for a general node in a graded mesh
that is also on a conductor or magnetic region boundary, it is helpful to start
with Helmholtz equation for the vector potential component A z• This is the
complex form of Poisson's equation and describes sinusoidal steady-state
behaviour, with ordinary Poisson's equation as the special zero-frequency
case. We now have the possibility of nonlinear permeability, or reluctivity v
(or strictly, with Helmholtz equation, constant J.1 or v that change from one
cell to the next), and the symmetrical form of Helmholtz equation in x, Y
coordinates is
:x [v ~~J + ~ [v ~;J = jrooA - Js (2.119)
where the current density is given the suffix s to denote an imposed current
(which we assume to be a real quantity for convenience) to distinguish it from
an induced current in the time-dependent system. Consider the four graded
cells shown in Fig. 2.8, each cell having specified dimensions hi etc., and
current density, conductivity and reluctivity. For ease of description the
compass point notation is used, e.g. ONE is the conductivity of the cell to the
North-East of the central node 0 under consideration. The most convenient
method of obtaining a finite-difference equation is to apply Stoke's theorem
II (curl A) . ds = 'A . dl (2.120)
to an auxiliary cell bdfh centred on node 0 [2.43]. Thus we require to perform

a surface integration of the curl form of eqn (2.119),
curIz (~IV
oA -:- OAJ .
oy - Jv ox = Js - jroaA (2.121)
over the cell bdfh, to give the line integral equation,

N E
; ( ~ oA
IV oy -
~ oA
JV
J
ox . dl = I -wI (Js - jroaA)
-5
dxdy (2.122)
where E =h/2, N =h212, W = h312, and S = h412. On ab the contribution

to the line integral is
N
Io ["iv oA
~ - "OAJ"
dy
jv ~ . j dy
vX
=- I
N
0
VNE oA
~
vX
I
x=E
dy (2.123)
Now oA / ox at x =E can be expanded as a Taylor series about point a, at which

A=A a ,
(2.124)
But
(2.125)
and
A =A _EoAI +~E202AI_0(E3) (2.126)

o a ox a 2 ox 2 a
Subtracting eqn (2.126) from eqn (2.125) we have
oA
ox
I = Al hi- A o - 0(E
a
2)
(2.127)
and substituting eqn (2.127) in eqn (2.124)
oA
ox
I x=E
= Al -
hi
A o - O(N) (2.128)
The approximate value of the line integral along ab is therefore -NVNE(A I -

A o}/hi with an error of 0(N2), and, following similar arguments, the com-
plete line integral of eqn (2.122) is given by
where
(2.129)
aZ = EVNE +h WVNW (2.130)

z
NVNW+ Svsw
a3 = (2.131)
h3
a4 = EVSE +h Wvsw (2.132)

4
The error is O(K z), where K is the larger of the half mesh dimensions N, S,
E and W.
The surface integral on the right-hand side of eqn (2.122) is the sum of the
integral ov.er each quadrant of Fig. 2.8. Consider the first quadrant:
NE NE
ff (Is - jc.ooA) dxdy = NEJNE - jc.o0NE ffA dxdy (2.133)
00 00
A double Taylor expansion about node 0 yields
A =Ao+x aA
ax I +y aAay I +O(K)z
0 0 (2.134)
so that
NE
rr 1 aA
JJAdxdy=NEAo+-NEz- I +-NzE-
1 aA I +O(K 4
) (2.135)
00 2 ax 0 2 ay 0
If we ignore terms O(K 3 ), bearing in mind that the approximation to the

line integral has an error O(K z), the complete surface integral has the form
10 - jc.oQo, so that eqn (2.122) finally becomes
A - (aldl + azdz + a3d3 + a4d4) + 10
(2.136)
_0 - (al + az + a3 + a4) + jc.oQo
where 10 is the total imposed current associated with node 0, i.e.
10 = NEJNE + NWJNW + SEJSE + SWJsw (2.137)
and similarly,
Qo = NEo NE + NWo NW + SEOSE + SWo sw (2.138)
The simple notation of eqn (2.136) can easily be generalized to apply to any
node (i, j), with the cell parameters being given the same address as the node
in the bottom left-hand corner of the cell, e.g. the cell in the second quadrant
of Fig. 2.8, at present labelled NW, is now cell (i - I, j) and VNW = Vi_I,j etc.
The important point is that eqn (2.136) is completely general and can be
applied to a node in any position including all types of internal boundary. It
is also highly suitable, in its static or diffusion form, for nonlinear problems
where the reluctivity will be different in each cell, and a function of the average
values of the derivatives of A in the cell after each scan of the iterative
solution, e.g. since
(2.139)
we have, for cell (i, j)
B2= {~[Ai+l.j - Ai.jl + ~ [A i + l •j + 1 - Ai.j+I]}2

av 2 t1x J 2 t1x
+ H(Ai.j+~; Ai.j] + ~ (Ai+l.j+~; Ai+l,jJY (2.140)
This is why it was suggested earlier that one useful way of storing the
magnetization details of a magnetic material is in the form of a B 2Iv file.
Equation (2.136) can be solved by successive over-relaxation (SOR) as
described in Section 2.6.4. However, for a nonlinear solution, solved in terms
of A, it is essential to under-relax the cell reluctivity v obtained from the value
of B 2, so that, for cell (i, j),
V(k+l)
'.J
= V(k)
'.J
+ w(v _ V(k»)
'.J (2.141)
where w is less than one (typically 0.2, but problem dependent).
Clearly, for a ferromagnetic region, a reluctivity map in the form of a
separate array of cell values is required. However, where J and 0 are constant
in large conducting regions two more full arrays would be extremely wasteful
in space and so some form of rationalization is necessary, even to the extent
of listing special versions of eqn (2.136) applicable to each region and boun-
dary node type. In the extreme case of a node in the interior of a region of
constant conductivity, reluctivity, and current density, that has been meshed
with a grid of simple square cells, eqn (2.136) reduces to
AI + A + A + A + 1l01l,1
A o = --------------
2 3 4 0 (2.142)
4 + jrolloll,Qo
where 10 = Jt1x 2 and Qo = ot1x 2 • It can be seen that eqn (2.106) will reduce to
the same equation if t1y = t1x and 0 = O. The SOR solution of a linear complex
equation such as eqn (2.142), or eqn (2.136) with constant values of n, ideally
involves a complex relaxation factor. A near optimum value of the latter can
be derived during the solution using a modified version of the Carre method
[2.9; 2.42; 2.43] mentioned in Section 2.6.4.
It is worth noting that in the alternative two-dimensional problem where the
magnetic field is perpendicular to the x,y plane, i.e. we have Hz only, the
diffusion equation in H is
~ (p _aHJ + ~ (p _aHJ = jrollH (2.143)

ax ax ay ay
and is analogous to eqn (2.119) with H in place of A, resistivity p in place
of reluctivity v, and permeability Il in place of conductivity 0. The only
difference is the absence of the current density term that appears in eqn
(2.119). Thus, since entirely mathematical arguments were used in the deriva-
Ow
3 :0 1
Fig. 2.9 Small air gaps inserted into a rectangular grid
tion of eqn (2.136), a finite-difference approximation to eqn (2.143) has the

same form as eqn (2.136) but with 10 zero, (J NE etc. in eqn (2.138) for Qo
replaced by IlNE etc., and similarly VNE etc. in eqns (2.129) to (2.132) for the
a coefficients replaced by PNE etc.
A very useful feature of the way in which eqn (2.136) was derived is that
the presence of very small air gaps between highly magnetic surfaces can be
incorporated in the line integral formulation. To actually mesh a very small
air gap is expensive in nodes because even the cell dimension parallel to the
gap is restricted by the aspect ratio of the cell, which should not be greater
than 10 to 1. Suppose the gap is 0.5 mm and we require at least two cells to
straddle it. Typical cell dimensions then become 0.25 mm x 2 mm, say, which
in an extensive air gap, will result in a large number of cells.
Inclusion of the very small air gaps illustrated in Fig. 2.9 can be achieved
by adding their mmf contributions to the line integral of eqn (2.122). In air,
we have
Hx = _~.JOAJ and H y = - ~ [OAJ (2.144)

Ilo Lay Ilo ax
If the relative permeability is sufficiently high for the flux to cross the small
gap perpendicular to the iron surfaces, then the contribution of gap 0E' for
example, is simply
0EHy = 0E [_ OAJ : : : 0E A o - AI (2.145)

Ilo ax Ilo hI
We note that points on opposite sides of the air gap both share the same node
number under this assumption. The total contribution of all four air gaps to
the mmf, bearing in mind that the line integration is anti-clockwise, is
i+ l,j
Fig. 2.10 A node in a polar grid with the surrounding cells
From the form of eqn (2.136), we thus see that a term of the form OI(J.l oh)
must be added to each a coefficient, so that, for example, ~3 = a3 + owl
(J.l Oh 3), and the new finite-difference equation is identical in form to eqn
(2.136) but with the new coefficients ~j in place of aj(i = 1-4). This is a
general form of the. equation and does not imply that gaps must exist in all
four branches at once.
Stoll [2.44J has applied the above technique to the problem of calculating
the transverse force on a magnetic filler block in a flexibility slot in the solid
steel pole face of a two-pole turbogenerator rotor where the likely clearance
gaps are less than 0.25 mm, but not uniform.
The alternative two-dimensional geometry is the polar coordinate system
illustrated in Fig. 2.10. The Helmholtz equation equivalent to eqn (2.121) now
becomes
" v oA "OAJ .
curIz [ ar~ 09 - aava;: = Js - JrooA (2.146)
The diffusion form has ooAlot in place of jrooA.

Applying Stoke's theorem in a similar way to that for the rectangular
system, Wiak [2.55J shows that the finite-difference equation is
(a 1 + az + a3 + a 4)A j,j - (a1Aj,j-l + azAj+l,j + a3 A j,j+l + a4 A j_l,j)

=i
r·
[(hj+l£\9jJNE + hj+l£\9j+1JNW + hj£\9j+IJsw + hJi.9j JSE )
- (hj+l£\9jONE + hj+l£\9j+10NW + hj£\9j+10SW + hj£\9joSE)jroAj,j]
(2.147)
where
(hi+1VNE + hivSE) (2.148)

at=
2ri~eJ
a2 -_ (~eJ+IVNW+ ~eJVNE)(ri+
_~
-!h i + l )
-'----'--=---'-'_--=---,-,-:..c..
(2.149)
2h i + 1
(hivsw + hi+tVNW )
a3 = ----...;..;;..-'--- (2.150)
2ri~eJ+l
- -! hi)
a4 = (~eJ+tVSW + ~eJVSE)(ri
2h
(2.151)
i
Equation (2.147) has been presented in this way, with Ai,J on both sides,
so that the modification to the diffusion equation is easily accomplished
by replacing jroA on the right-hand side by aAlat in the desired numerical
form.
The components of flux density in a given cell can be obtained by averaging
the derivatives of A on opposite sides of the cell, e.g. in the North West cell
in Fig. 2.10 (cell (i,j) bearing in mind the directions of increasing i and j) we
have
B -! aA :::::! (Ai+1'J+l - Ai+l,J + Ai,J+l - Ai,JI (2.152)

, - r ae 2 ri+l~eJ+l ri~eJ+t j
Wiak and Pelikant [2.57] interestingly solve the static Poissonian version of
eqn (2.147), Le. without the last induced eddy-current term, for a switched
reluctance motor. However, the earliest and major application of the above
forms of nonlinear Poisson's equation (modified versions of eqns (2.136)
and (2.147» was to the analysis of d.c. machines by Erdelyi and Fuchs [2.16;
2.18]. Although computer programs of this complexity are unlikely to be
written today, in view of the present widespread availability of commercial
and in-house general purpose finite-element packages, one interesting feature
of the work in the 1970s was the development of ways of improving the con-
vergence of steady-state iterative methods such as SOR. In this respect, the
integral correction method of De la Vallee Poussin [2.13], being based on
Ampere's law (eqn (1.27», has been found to be helpful.
Consider a contour such as that shown dotted in Fig. 2.11 which is close
to the boundary of the problem and bisects a series of mesh links that are
locally perpendicular to the boundaries. The contour encloses a known total
current I. Let ~n denote the length of these perpendicular links, that may vary
along the contour, and M the tangential distances between adjacent nodes.
Since B = curl A or H = v curl A, a numerical form of Ampere's law applied
to the contour becomes, in terms of Ax:
~ v(A i - Ao)av [~~J = I (2.153)
where ~n will be constant on the straight sections of the contour and

(Ai - Ao)av represents the average difference between the values Ai on the
yt1----------,
I I I L1t i
z
l-_----W-----
(i;.----------------.. . .
I
--~ .
x
Fig. 2.11 Integration contour for integral correction method of improving convergence
corners of a given cell (of reluctance v) inside the contour and the values A o
on the corners outside the contour. However, at a particular stage in the
iterative process eqn (2.153) will not hold, because the solution has not con-
verged. Let a constant value M be added to all nodes inside the contour, such
that eqn (2.153) is satisfied, Le.
~v(Aj + M - Ao)av [:~J = I (2.154)
Thus, solving for the constant M, we have
[I - ~ v(A j - Ao)av [~J]

M= (2.155)
~v[:~J
Although this technique is not known to have been applied to steady-state
solutions of Helmholtz equation, where the vector potential component is
complex, there appears to be no reason why it should not be as effective as
it is for static solutions.
2.6.6 Two-dimensional transient or nonlinear steady-state problems

With time-stepping problems in two space dimensions it is almost essential to
employ an explicit method such as DuFort-Frankel, or a semi-implicit method
that reduces or simplifies what would otherwise be a full matrix solution for
a large number of nodes at each time step. If jroA on the right-hand side of
eqn (2.147) is replaced by
ij,k+l
i-lj,k
i,i,k-l
Fig. 2.12 Finite-difference molecule in two space dimensions (i, j) and time (k)
aA Ai,j,k+1 - Ai,i,k
- "" --""'-'-'-.:....:...._....;.:,,:.= (2.156)
at M
and Ai,j,k is then replaced on both sides by (Ai,i,k+ 1 + A i,i,k-l)l2, we have the
DuFort-Frankel form of the diffusion equation fer A z in polar coordinates.
Omitting the imposed current density term, and writing
Qi,i = t ri(hi+1 ~ejONE + h i+1~ej+1 0NW + hi~ei+l 0SW + hi~eioSE)
(2.157)
for brevity, eqn (2.147) becomes
Ai,i,k+l [al + a2 + a3 + a4 + ~:] = 2a A i,j_I,k + 2a2Ai+I,i,k + 2a3Ai,i+l.k

1
+ 2a4Ai-l,j,k - Ai,i,k-I
[al + a2 + a3 + a4 - Qi'i]
~t (2.158)
The difference molecule is shown in Fig. 2.12.

Extending the work of Richmyer and Morton [2.35], who examined the
stability of three-time-Iayer difference schemes using the method of energy
inequality, Wiak [2.55J shows that the DuFort-Frankel scheme must satisfy
the condition
2M(~r2 + r2~(2)
1+ r2~r2~e2o (Vmin - v max ) ~0 (2.159)
where the radial mesh length, ~r, and angular displacement, ~e, are now
assumed to be constants. With strong nonlinearity, Vmax ~ Vmim so that
eqn (2.159) reduces to a condition on the time step,
r2~r2~e2o
M ~ (2.160)
"" 2(~r2 + r2~(2)vmax
Wiak [2.55] also shows that eqn (2.160) holds in the nonlinear case, and goes
on to solve the problem of the eddy currents induced in a ferromagnetic rotor
screen of a synchronous generator with a superconducting d.c. field winding
when the rotor oscillates (hunts) about synchronous speed. The onset of
numerical instability is shown as the time step increases above the value of
33 J!S set by eqn(2.160) for the parameters involved.
The above example actually solves the DuFort-Frankel algorithm inside the
conducting screen but uses analytical solutions outside, in terms of the Fourier
series that result from the normal separation of variables method applied to
Laplace's equation. The analytical and numerical solutions are joined at the
boundaries of the conducting region. A more general concept is that of linking
the numerical diffusion algorithm with one suitable for Laplace's equation,
e.g. SOR. This boundary matching can be illustrated as follows. At each new
time step the nodal values of A z on the surface of the conducting region are
used as Dirichlet boundary conditions for the adjacent air space. The updated
nodal values within this Laplacian region are found by performing a few SOR
iterations. For example, in the air gap of an electrical machine about five
iterations are sufficient, not only because of the limited air gap length but also
because the change in the interface values from one time step to the next is
small. The row of nodal values just inside the air region and parallel to the
interface are then used as the boundary values for the next DuFort-Frankel
step in the conducting region, so that the two regions overlap by one layer
of nodes. In the process of moving from conductor to air the normal com-
ponents of B, in terms of the surface nodal values of A z , are supplied.
Returning from air to conductor, the tangential components of B are effec-
tively specified by the values of A z at each pair of nodes perpendicular to the
air side of the interface. Jack and Stoll [2.22] have used this technique to find
the negative sequence induced currents and ohmic losses in the solid rotor of
a turbogenerator.
A popular semi-implicit method is the Alternating-Direction Implicit
method (ADI) first proposed by Peaceman and Rachford [2.31], and
examined in great detail in the book by Wachspress [2.52] for elliptic
equations. For the purpose of ease of demonstration we will take the simple
implicit form of the diffusion version of eqn (2.105) in x,y coordinates for a
uniform rectangular cell size, tu x ~y, and constant permeability and con-
ductivity, i.e.
1
(tu)2 (Ai+I,j,k+1 - 2A i,j,k+1 + Ai-I,j,k+l)
1 OJ!
+--(A"
(~y)2 Ik + 1-2A"k
',J+, ',J, + I+A"',J- Ik
, + l)=-(A"k
~t ',J, + I-A'k)
',J,
(2.161)
The ADI algorithm moves forward in time with a repetitive two-step cycle.
In moving from time k to k + 1, eqn (2.161) becomes half explicit in the y
direction, to give
1
(~XV (Ai+I,j,k+1 - 2A i,j,k+1 + Ai-I,j,k+I)
1 a~
+--(A--+
(~y)2 , -2A-I,J.-k+
I , Jlk - - Ik)
A I,J- , =-(A--
I1t 'oJ, k+ I-A--k)
I,J,
(2.162)
whereas, in moving from time k + 1 to k + 2, eqn (2.161) becomes half explicit
in the x direction,
1
(~)2 (Ai+I,j,k+1 - 2A i•j ,k+1 + Ai-I,j,k+I)
1
+- - (A-I,J+,
(~y)2 - I k+ 2 - 2A-I,J,- k+ 2 + A-I,J-
- I , k+ 2) =-a~
~t (A-I,J.-k+ 2 - A-I,J,-k+ I)
(2.163)
At each time step, therefore, sets of one-dimensional implicit equations must
be solved, each set having a simple tri-diagonal coefficient matrix. In the linear
case, the process is simple and efficient because the coefficient matrix remains
constant throughout and can be decomposed into the product of lower (L) and
upper (V) triangular matrices, where, because of symmetry, U is equal to the
transpose of L. L can be found by the Choleski method for the larger of the
row or column dimensions. Brief details are given by Stoll [2.43], including
the use of ADI for steady-state complex solutions. The main difficulty with
the ADI method is that its successful performance is very problem dependent.
This follows from the substantial explicit component. Thus it is worth inves-
tigating any alternative that increases the implicit content, but without going
to the computational burden of the two-dimensional form of the simple
implicit equation (eqn (2.161» or the Crank-Nicolson algorithm. One possi-
bility is the interesting strongly-implicit iterative method of Stone [2.46],
outlined in some detail by Stoll [2.43].
2.6.7 Three-dimensional solution of Laplacian and Poissonian problems

Laplacian problems can be solved in terms of a magnetic scalar potential such
that H = - grad V. The magnetic flux density satisfies div B = 0, and so,
div (~H) = div (~grad V) = 0 (2.164)
Applying Gauss' theorem (eqn (1.22» to the small volume surrounding node
oof Fig. 2.13, where the faces of the volume bisect the mesh links, the volume
integral of the divergence becomes the surface integral of the vector ~ grad V,
so that
~ ~ grad V . ds = 0 (2.165)
The derivation of a three-dimensional finite-difference equation now follows

a similar process to that used for the development of the Helmholtz (or
diffusion) equation equivalent in two-dimensions, where a line integral in
outside (0)
3
south (S)
Fig. 2.13 Three-dimensional finite-difference molecule showing surface of integration
terms of A z was involved (see Section 2.6.5). Each of eight three-dimensional

cells, not fully shown in Fig. 2.13 but all sharing node 0, is defined by a
constant angular increment ~e but varying linear dimensions hi' h 3 , h s and
h6 in the radial and axial directions. For example, hi is not shown to avoid
confusing the diagram but is the mesh length between nodes 0 and 1 in the
positive radial direction. In the more general case, each cell can have a
different permeability, so that IlNEI stands for the permeability of the cell to
the North East and inside (Le. decreasing z) of node O. Leurs and Stoll [2.27]
give the finite-difference equation that results from approximating the surface
integral over the auxiliary cell of Fig. 2.13
VI-VO [
4h
1
ro +
1]
2 hi ~e[hS(IlNEI + IlNWI) + h6(IlNEO + IlNWO)]
V2 -VO [
+ 4ro~e hs(hlllNWI + h 3 1lsWI) + h6(hlllNWO + h 3 Ilswo)]
-VO [ ro - 2
+ V34h J
I h 3 ~e[hS(llswI + IlSEl) + h6(Ilswo + IlSEO)]
3
V4 - Vo [
+ 4ro~e hs(htllNEI + h 3 IlSEI) + h6(htllNEo + h 3 IlSEO)]
+ Vs4h
- Vo ~e [hi[ro +
s
I]
:4 hi [ :4I]
(IlNWI + IlNEl) + h3 ro - h 3 (llswI + IlSEI) ]
6
+ V ;;'6 V
o
[hi + ~e ~O l hi] (IlNwO + IlNEO) + h 3 ~O -l h 3] (Ilswo + IlSEO)]
=0 (2.166)
The resulting symmetric set of linear equations can be solved by block
successive over-relaxation as an efficient alternative to normal SOR in three
dimensions. The unknowns that form a block are the potentials on an r, e
plane (z = constant). To minimize the band width, the nodes are numbered
in the circumferential direction first. Each block is solved by Cholesky decom-
position. In any block containing nodes that are all in air, the block matrix
only needs to be decomposed once. For example, with hi = h3 = I:1r, and
I.l. = I.l.o, eqn (2.166) reduces to
V ;0
~~ ~o + ~ I:1rJ l:1e (h s + h 6) + V~~;o ~o - ~ I:1rJ l:1e (h s + h6)
V2 - Vo V4 - Vo
+ 2r l:1e I:1r(h s + h6 ) + 2r l:1e I:1r(h s + h6 )
o o
V - V V - V
+ s h orol:1el:1r + 6h oro l:1e I:1r = 0
s 6
(2.167)
and the relative ease with which a set of equations of this form could be
handled is clear.
Suppose we now have a current sheet in the plane r = ro, carrying a surface
current of density
(2.168)
subject to the condition
diu K = ~ oKe + oK, = 0 (2.169)

r oe oz
The nodes in Fig. 2.13 can now be considered to be split by the infinitely thin
sheet as shown in Fig.2.14. Equation (2.165), with I.l. = I.l.o, must still be
satisfied over the auxiliary cell surface (not shown in Fig.2.14), so that
eqn (2.167) becomes
al(VI - Vo) + a3(V3 - Vo) + -!-a2[(V2 - Vo) + (Vi - vo)]
+ !a4[(V4 - Vo) + (v.; - Vo)]
+ ! as [( Vs - Vo) + (V; - Vo)]
+-!-a6[(V6 - Vo) + (V6 - Vo)] =0 (2.170)
where the a coefficients follow from eqn (2.167). The terms of the form
V/ - Vo can be found by applying Ampere's law to the vanishingly thin
rectangular paths formed by the corner nodes 0, i, i', 0'. For example, con-
sider the path 044' 0' in Fig. 2.14, for which ~ H . dl = I becomes
r ol:1eHe04 - r ol:1eHo04 "" Kz04rol:1e (2.171)
where the positive directions of the components of current sheet density
normal to the paths are shown in Fig. 2.14. Now He = -(l/r)(av/ae) so that
eqn (2.171) finally becomes
(V'; - Vo) = (V4
Vo) + rol:1eKz04
- (2.172)
and V.; - Vo can be eliminated from eqn (2.170). Similarly for the other
three paths we have
Fig. 2.14 Three-dimensional finite-difference molecule with split nodes for current sheet
insertion
(Vi - Vo) = (v2 - Yo) - rO~eKz02 (2.173)

(V; - VO) = (Vs - Yo) + hsKaos (2.174)
(V6 - Vo) = (V6 - Vo) - h6 KOO6 (2.175)
The only unknown primed term that remains in eqn (2.170) is (V3 - Vo). This
can be determined by considering the total current io flowing in the sheet in
one of the coordinate directions (either e or z) between node 0 and some
specified reference point where V = O. Thus Vo= Vo + io and
(2.176)
Thus the terms in the original finite-difference equation (2.167) can be simply
modified to produce the equation that applies to nodes on the surface of a
defined current sheet. We also note that, if it is inconvenient to specify both
K a and K z everywhere, one component can be generated from the other using
eqn (2.169).
If it is necessary to enter a current carrying region, Le. the current sheet
simplification is unacceptable, Carpenter's electric vector potential (T vector),
see for example [2.7], can be used. Here the magnetic field vector is defined
as
H =T - grad V (2.177)
where curl T = J to satisfy Ampere's law. The advantage of the T vector in
problems in which the current density vector is two-dimensional, or approx-
imately so, is that T can be represented in terms of only one component,
orthogonal to the current directions. Reference [2.8] shows an example of the
use of this technique. The T vector is also applicable to current sheets as an
alternative to the previous development entirely in terms of V [2.27].
The alternative for general current carrying Poissonian regions is a full
development in terms of A or H. For example, Sarma [2.37] gives finite-
difference equations in terms of Ax, A y and A~. However, it is unlikely that

complicated three-dimensional finite-difference approaches will be a produc-
tive development in situations where good finite-element software already
exists. Notwithstanding, some interesting and quite powerful uses have been
reported in the literature over the last three decades, e.g. nonlinear cartesian
Laplace in V [2.23], linear cartesian Poisson in A [2.12; 2.60; 2.61] (the first
reference being an example of the use of electrical networks in field solutions),
linear cartesian Helmholtz in H [2.45] with multiple zoning of resistivity.
3
Finite-Element and
Boundary-Element Methods
Ryszard Sikora
3.1 INTRODUCTION
Finite-elements and boundary-elements methods are widely applied in electro-
magnetic field analysis and synthesis [3.2; 3.3; 3.4; 3.5; 3.7; 3.10; 3.17; 3.22;
3.23; 3.25; 3.28; 3.29; 3.30; 3.31; 3.32; 3.33; 3.34; 3.35; 3.36; 3.38].
The development of computers made it possible to solve many problems
using numerical methods. The combination of the availability of computers
and the Finite-Elements Method (FEM) and Boundary-Elements Method
(BEM) contributed to establishing the methods and enabled them to be applied
to many engineering problems. In the past 30 years a large number of com-
puter programs based on FEM and BEM have been developed. Application
of FEM can be dated back to 1967 when O.C. Zienkiewicz used the FEM for
calculations of mechanical structures. The FEM was introduced by P.P.
Silvester and M.V.K. Chari for electrical engineering calculations [3.7; 3.30].
This method can be used in calculations of stationary and transient, linear and
nonlinear, isotropic and anisotropic, homogeneous and non-homogeneous
electromagnetic problems. FEM finds application in calculation of quantities
describing electromagnetic field distribution (e.g. E, H, D, B, J, A, T, V, Q)
and integral quantities (e.g., P, W, L, R). The method is quite useful for
computation in bounded regions, but is inconvenient when the region is
unbounded. There are several useful methods for electromagnetic field calcu-
lation in unbounded regions, e.g. the Infinite-Elements Method (lEM) [3.18;
3.19; 3.27].
Application of the boundary-element method can be dated back to 1978
when C.A. Brebbia published his book about BEM [3.5]. This method can
be used for electromagnetic field calculation in bounded and unbounded
regions in linear media.
3.2 WEIGHTED RESIDUALS [3.15; 3.35; 3.38; 3.40)

The weighted residual method can be used as a basis for the finite-elements
method and boundary-elements method [3.15]. It is helpful in unifying
software design for both FEM and BEM. Consider a linear, operational
equation:
L (uo) =0 in space Q (Fig. 3.1) (3.1)
Weighted Residuals 79
Fig. 3.1 Bounded region
with boundary conditions:

S(uo) =s (3.2)
G(uo) = g (3.3)
where:
r = r l + r2
L N-order differential operator,
G N - I-order differential operator,
S N - 2-order differential operator,
Uo unknown function.
Equations (3.2) define the essential and natural boundary conditions. In this
section differential equations of the second order (N = 2) will only be taken
into account so the essential boundary conditions mean Dirichlet and the
natural boundary conditions mean Neumann conditions.
The exact solution Uo is approximated by a complete series of functions
[<pJT,j = 1,2, ... ,
(3.4)
where:
u an approximating function,
ai unknown coefficients,
<Pi linearly independent functions.
An error inside the domain Q is determined by a relation:
E = L(u) - L(uo) (3.5)
Similarly on the rI border
(3.6)
and on the r2 border
(3.7)
80 Finite-Element and Boundary-Element Methods
Depending on the type of approximating solution of eqns (3.1)-(3.3), part

of the errors is minimized and part is neglected. Let us assume that an error
is equal to zero on the r border. This assumption is used in the spatial
methods. The error is minimized by the formula
(E, w) == (L(u), w) == I L(u)wdO = 0

o
(3.8)
where:
w = ~.'II. + ~2'112 + ... + ~j'llj + .. , + ~m'llm, (3.9)
~j any coefficient,
'IIj linearly independent functions.
The application of the weighted residual method depends on the selec-
tion of appropriate function series ['IIa T • Some possible choices are discussed
below which leads to the following particular methods.
Point collocation method [3.8J

If
'IIj = OJ
then
(L(u),o;) =0 for i= 1,2, ... ,m (3.8a)
where OJ is the Dirac delta function in point i.
A set of equations (3.8a) may be over-determined when m > n (n was
defined by eqn (3.4».
Bubnov-Galerkin method [3.14J

It is very easy to introduce FEM on the basis of the Bubnov-Galerkin method.
In expressions (3.4) and (3.9) we assume that n = m. Thus
u = u.(j>. + U2(j>2 + + Uj(j>j + + Um(j>m (3.10)
w = ~1(j>1 + ~2(j>2 + + Pj(j>j + + ~m(j>m (3.11)
*0
= {= 0
in OJ
If (j>j in 0" OJ
Where OJ is a finite element, we obtain a numerical implementation of the

Bubnov-Galerkin method which is the FEM.
The least-squares method [3.8, 3.14J

This method is based on minimizing the square of error E2 = (L(u) - L(UO»2,
which is equivalent to the assumption that 'IIj = L«j>j).
If the operator L (3.1) is symmetric, integration by parts is usually used,
and instead of the scalar product (L(u), w> the following relation is assumed
(L(u), w) = - (D(u), D( w) + (G(u), S( w)r, (3.12)
Weighted Residuals 81
where
(G(u), S( w) >r == J G(u)S( w) dr (3.13)

r
For the Laplace's operator (N = 2, L = V2) eqn (3.12) is similar to Green's

formula
J V2uwdQ = - J gradugradwdQ + J au dr (3.14)

Cl Cl r on
L is an N-order operator and D is of order N - 1. In this case it is possible
to use basic functions [q>;]T belonging to the space C N - 2(Q). For the case
(N = 2) functions [q>;]T are continuous and their derivatives are disconti-
nuous. Moreover, if the operator L is positive definite the equations obtained
from the Bubnov-Galerkin method are the same as equations obtained from
Ritz's method.
The boundary methods belong to the next group. They are based on the
assumption that functions [q>;]T or [WilT satisfy eqn (3.1). According to this
£ = 0 whereas £} and £2 are minimized. Let us present boundary methods for
the cases where L is a symmetric operator.
Boundary collocation method [3.8J

If functions [q>;]T satisfy eqn (3.1) but do not fulfil the boundary conditions,
the error is minimized on the boundary
(£., w>r. == (S(u) - s, G( w) )r l =0 (3.15)
(£2' w>r2 == (G(u) - g, S( w) )r2 = 0 (3.16)
Trefftz method [3.8}

If functions'll; satisfy eqn (3.1), integrating by parts the relationship (3.8), we
obtain
(L(u), w) == (u, L*( w» + (S( w), G( w)>r = (S(u), G*( w)r = 0
(3.8c)
where L *, G * are the operators adjoint to L, G.
Substituting eqn (3.1) into eqn (3.8c) the following boundary equation is
obtained
(S(w),G(u»r = (S(u),G*(w)r (3.17)
Method based on the singular solution

Assuming that the weight functions W; are solutions of the equation
L*( w) = 0; (3.18)
in the domain Q; the following equation is obtained
u; + (S(w), G(w»r = (S(u), G*(w)r (3.19)
and on the border r surrounding domain OJ

CjUj + (S( w), G(u»r = (S(u), G*( w)>r (3.20)
Equation (3.20) is the basis of the BEM.
3.3 THE FINITE-ELEMENT APPROACH

[3.4; 3.26; 3.31; 3.38]
The error in the Bubnov-Galerkin (Ritz) method can be minimized by:
• using a large number (m) of elements in the approximating formula

(3.10);
• dividing the domain 0 into many subdomains OJ (elements).
In the latter case a simple approximating function (small m) is applied

within the element OJ. The classical Bubnov-Galerkin (Ritz) method is based
on the first approach. More complicated approximating functions (larger m)
must be used for more exact solutions. Complicated functions need sophis-
ticated programming and time-consuming calculations. The complexity of the
functions depends on the shape of the domain border.
FEM is based on the second approach. The FEM may be implemented
in several ways. For two-dimensional problems the simplest subdomain
(element) is a triangle with three vertices (nodes). The required function u e in
this kind of element is linear with respect to coordinates. Derivatives of u e
are constant inside element ([<p;]T are of CO class). Between the elements, the
functions u e are continuous. In order to have a nearly exact solution (u "" uo)
and fairly smooth derivatives of u we have to use a sufficiently dense tri-
angular mesh. Alternative FEM formulations will be presented in Section 3.4.
FEM will be presented as a subdomain technique based on the Ritz's method.
Consider the operational equation
L(u) =! (3.21)
which has a different right-hand side from eqn (3.1). Operator L is positive
definite. The solution of eqn (3.21) is provided by minimizing the functional
cI>(u). The shape of functional cI>(u) depends on: the operator L, ! and the
boundary conditions. A few functionals will be shown later. In applying FEM,
the domain, in which functional cI>(u) is minimized, is divided into subdomains
OJ (finite elements). This is shown in Fig. 3.2.
Elements are defined by nodes (vertices) (i,j, k). For simplicity only tri-
angular elements will be considered for two-dimensional problems. Inside any
triangular element OJ the solution u e fulfils the following relationship
u' (x, y) ~ [N,(x, y), 1'.j(x, y), N, (x, y) I [~] ~ [N]T [u l' (3.22)
The Finite-Element Approach 83
Fig. 3.2 Division of domain into elements
where:
[N]T row vector of the shape functions,
[u]e column vector,
set of the function u with values at nodes i, j,
k, respectively,
Ni(x,Y), ~(x,y), Nk(x,Y) must be unity at the ith, jth, kth nodes res-
pectively and zero at all other nodes (e.g.
=
Ni(Xi,Y;) 1, Ni(xj,Yj) 0). =
The functional (u) is minimized with respect to values of the function u in
all the nodes (vertices) of domain n (together with border n,
which means
with respect to a vector:
[u] = (3.23)
u,
where r is the total number of nodes after division of domain n into elements.
In order to minimize the functional (u), (u) must be differentiated with
respect to column vector [u] and equated to zero
a<l>
au!
a<l>
-
a<l> aU2
=0 (3.24)
a[u] =
a<l>
au,
Taking into account that

t
 = ~ <l>e (3.25)
e=l
where:
 total functional in domain a,
<l>e functional in subdomain a e,
e element number,
t total number of elements,
we obtain for any equation belonging to a set (3.24)
(3.26)
For any element, where we take into account an elliptical equation, we obtain
a<l>e
- = [h]e[u]e + [s]e (3.27)
arule
or
a<l>e
aUi
a<l>e
(3.27a)
au}
a<l>e
aUk
where:
[h) stiffness matrix,
[8] load vector.
Finally the set of equations transforms to
0<1>
a[u] = [H] [u] + [s] = [0] (3.24a)
where
Sl
H ll HI,
82
[8] = [H] = H il H ii Hi} Hi' (3.24b)
H'I H rr
S,
Ui
Yi
Fig. 3.3 Triangular elements in cartesian coordinates
I I
Hij = ~hij,. s; = ~Sr (3.24c)

e e
Vector [u] is a solution of equation (3.24) or (3.24a). Inside the triangular

elements (Fig. 3.3) function u is linear with respect to coordinates X,y:
ue{x,y) = (I, + (l2X + (l3Y (3.28)
At the nodes (vertices of triangles) function u satisfies the following set of
equations
u; = (I, + (l2X; + (l3Y;
Uj = (I, + (l2Xj + (l3Yj (3.29)
Uk = (I, + (l2Xk + (l3Yk
or in matrix notation
(3.29a)
or
[u] = [a] [(I] (3.29b)
Solving eqn (3.29) we obtain
(1\ =-2.6.I (a·u· + a·u· + OkUk)

1/ II
1
(12 =-2a (b·u·
./ + b·u·J + bkUk)
I I
(3.30)
1
(13 = -2.6. (c·u. + c·u·
I J J
I
+ CkUk)
where
2A ~ de{: (3.31)
11 is the area of a triangle i, j, k, and

a; = XjYk - XkYj
b; = Yj - Yk = Yjk (3.32)
Ci = Xk - Xj = Xkj
Coefficients aj' bj , ••• are calculated by cyclic permutation of suffices i,j, k:

aj = XkYi - XiYk
bj = Yk - Yi (3.33)
Cj = Xi - Xb'"
Substituting eqn (3.30) into eqn (3.28) we obtain eqn (3.22), where
I
Nm = 211 (am + bmx + cmy) m = i,j, k (3.34)
The method is applied to the solution of a two-dimensional problem, described

by Helmholtz's equation:
ap
-- (au]
---pa -(au]
+ k u2= f (3.35)
ax x iJx iJy y oy
If Px =Py = 1 the first part of eqn (3.35) is simplified to
2 2
a [au] a [au] au au (3.36)
ax ax + ay ay = ax + ay2 = flu
2
Equation (3.35) is valid in anisotropic (Px Py) and non-homogeneous *"
(PAx, Y), pix, y»
media. In isotropic and homogeneous media, function
u(x,Y) satisfies the simplier eqn(3.36). Coefficients PX,Py are physical
parameters e.g. En Ey , !lx' !ly, or an 0y, according to the prescribed field
(electrostatic, magnetostatic, or electromagnetic). Coefficients Px and Py must
be differentiable whereas k 2 and f must be continuous. The shape of the
function depends on a differential equation and boundary conditions. For
Dirichlet's conditions
ulr = u(B) = Uo (3.37a)
where B is any point on the border r, the functional is as follows
<1l(u) = I (::Y (:;y

[Px + Py + k 2u2 - 2fU] dxdy. (3.38)
Minimizing the functional (3.38) is equivalent to a boundary value problem

(3.37a) for eqn (3.36). For Neumann's conditions
du
dn r
I = <p(B) = -<Po (3.37b)
The following functional is now obtained
4>(u) = 1[px(~:r + py(~;r + 2

k u
2
- 2f U] dxdy + 2! <Po udr
(3.39)
For Poisson's or Laplace's equations in functional (3.38) or (3.39) the relevant
components (k 2u 2 ,2fu) are neglected.
Equation (3.24a) is resolved in the following manner. At first the functional
must be differentiated with respect to a nodal value of Um
d4>e
-=2 J [PdUe
x -d- [due]
- dUe -
+p - d [due]
- dUe- fdue]
+k 2u e- - dxdy
dUm 0' dX dUm dX y dY dUm dY dUm dUm
r
+ 2 J <po;-dr,
dUe
m = i,j, k (3.40)
r' lJU m
where re is the boundary of the element ae •

dUe. [dN;
dX = dX'
d~ dNk]
ax' dX
[U;] 1 [
~ = 2.:\ b;, b j,
] [ ]e
bk U
dUe
dY =
[dN; d~ dNk] [U;]
dY' dy 'dy ~ = 2.:\
1 [
C;, Cj' Ck
] [ ]e
U
(3.41)
d~m (dd:eJ =:~ d~m (~:eJ =;~

dUe 1
dUm = Nm = 2.:\ (am + bmx + cmy)
eqn (3.40) becomes
:=~ = 21. [~~ 2 bmPxlb;, bj , bkHu]e + 4~2 cmpAc;, Cj' cd [u]e +
+ k 2 [N]T[u]eNm - fNmJ dxdy + 2 J <PoNm dr

r'
(3.42a)
If domain isdiscretized into a large number of elements ae, then PX,Py
a
and f may be assumed constant inside ae • Additionally taking into account
that
f N;dxdY = f Ajdxdy = f Nkdxdy = ~a (3.42b)

0' 0' 0'
we obtain
a~e I
aU = 2a [Pxb;b m + Pyc;cm, Pxbjbm + PyCjCm,Pxbkbm + PyCkCm] [U] e
m
+ f k2[N]T[U]NmdxdY-~Af+2r <poNmdr e
(3.43)
0' r'
or
a~e
au;
a~e
(3.44)
aUj
a~e
aUk
where, for example
h-. = Pxb;bj + PxC;Cj

'J 2a +
2 r keN-N dx dY
'J
(3.4S)
0'
2
f <poNmdr
qlo
S~ = -3 af + 2
r'
m = i, j, k (3.4Sa)
For the full a domain the following system of equations is obtained
a~ I a~e
~[
V U,
.] = e=1
~ ~[v U,
.] = [H][u] + [S] = [0] (3.46)
Inside aa function [u] satisfies the equation

[H][u] + [S] = [0] (3.47)
To illustrate the method mentioned above the problem shown in Fig. 3.4
will be solved. After discretization of a, the total number of nodes is equal
to S. In Fig. 3.Sa-e the procedure of formulating eqn (3.46) is shown. Figures
3.Sa, b, c and d present equations for successive elements '1 e• The first
element is defined by nodes 1,2,3. According to this matrices [H] I, [u] I and
[8] I were formed. Non-zero coefficients are marked by lines in cells. Empty
cells mean zero coefficients (Fig. 3.Sa). Using this procedure eqns (3.46) are
formed for elements 2, 3 and 4 (Fig. 3.Sb, c and d). In the next stage the global
matrix is written (Fig. 3.Se). It is easy to observe (Fig. 3.Se) that the global
stiffness matrix [H] is symmetric. Only nodes 1 and 4,2 and S are not directly
connected in this example (Fig. 3.4). Thus h l4 = h41 = 0 and h2S = hS2 = 0
(Fig.3.Se). If a domain were divided into many elements (e.g. Fig. 3.6) the
matrix [H] would include more coefficients equal to zero because only a few
].
Fig. 3.4 Division of domain into four triangular elements
first element four element

[H]' [u]' +[S]' =[0] [H]4[U]4+[S]4=[0]
1 234 5
1 1'\ 1'\ 1'\ 10- t-- I--
2 1'\ 1'\ r'\.

3 1'\ r'\. r'\. x + ={O} 10- I-- I-- x + ={O}
4
5 l- I- I-
a global space d
second element
[H]2[u]2 +[S]2 =[0] ·[HJ(u]+[S]=[O]
& e 5 e
[H]=1:JH] [S)=1:J SI
~ '\ ~ l- 0
/ / V f'\. )( / 0
/ / V x + ={O} !-"or )( x + = 0
/ / V /11 0
"- 0
b e
third element
[HJ3[U]3+[5]3=[0]
1--+-+++1-+-11-1 x + ={O}
c
Fig. 3.5 Assembling of FEM equations for the example from Fig. 3.4
1 2 3 4 5 6 7 8 9 10 11 12
1 x >< X >(
2 )( )( 1')(
>< X
3 X )( X XX
4 1)< X X
5 >< X >< ~X
6 >< )< )( ')( ')( X )<
2 3 4 7 )< L>< )( x x )< X
8 X X
>< X X
5 6 7 8 9 X x )(
10 X LX )( x x
11 x x >< XX
9 10 11 12
12 XX X~
Fig. 3.6 12-node mesh and its [H] matrix
nodes are connected directly. Thus in this case the matrix becomes a sparse
matrix. A sparse matrix may be banded if non-zero coefficients are placed in
a band. The band width is equal to the maximum difference in number of
nodes plus one. Thus in the problem presented in Fig. 3.6, the band width is
equal to 6 (e.g. 1st-element 6 - I + I =6, 6th-element 12 - 7 + I =6). The
matrix [H] is symmetric and banded. Only the upper triangular part of the
matrix, including the main diagonal, needs to be stored in a computer
memory.
3.4 ISOPARAMETRIC ELEMENTS (3.38; 3.40)

The simple three node triangular element was described in Section 3.3. Some-
times the boundary r of a domain is strongly curvilinear. In these cases an
accurate representation of r by triangular elements requires a very dense mesh.
A problem like this may be solved by using curvilinear elements. It is then
possible to obtain accurate representation of the boundary r by a small num-
ber of elements. In this section four elements will be considered, Le. (Fig. 3.7):
• four-nodal quadrilateral,
• six-nodal isoparametric triangle,
• six-nodal rectilinear triangle,
• eight-nodal isoparametric quadrilateral.
Within a three-node triangular element, the solution is given by three terms

(3.4). Generally a number of function terms is equal to a number of element
nodes. Thus for a four-nodal element
Isoparametric Elements 91
., -r
k
.J -\
j
fournodal sixnodal
quadrilateral isoparametric triangle
sixnodal eightnodal
rectilinear triangle isoparametric quadrilateral
Fig. 3.7 Elements including more than three nodes
(3.48)
and for six nodes
u = al + a2x + a3Y + a 4xy + asx 2 + a6y 2 (3.49)
It is very convenient to use local normalized coordinates (n, ~) in deriving
the isoparametric stiffness matrix. To achieve compatibility with Gaussian's
quadrature, coordinates (n,~) are defined in the interval (-1,1). For the two-
dimensional case, application of local coordinates leads to the equation
(3.50)
The local coordinates (n,~) of a point in the rectilinear quadrilateral

(Fig.3.8a) correspond to the point (x, y) in a curvilinear quadrilateral
(Fig. 3.8b) defined by the relationships
x(n, E,) = (N]T(x]e (3.51)
Y(n, E,)= [N]T[y]e
where
-1,1 1,1
o
-1,-1 ....... ...1 1,-1
<D
(a) (x"y,) <D (b)
Fig. 3.8 Four-nodal isoparametric elements
Yi]
[y]' -
[~
For the quadrilateral (Fig. 3.8) the shape functions are defined by
(3.52)
where
111 = 114 = - I
Tl2 = 113 = 1
(3.53)
~1 = ~2 = -1
~3 = ~4 = 1
On the sides of elements (TI = ± 1 or ~ = ± 1) shape functions Nm(TI, ~) are
linear. They satisfy the following conditions:
M
~ Nm(TI,~) = 1 (3.54)
m=1
Global coordinates x, yare explicitly defined by local coordinates TI,~. The

inverse is not always true. Fortunately as a rule there is no need to calculate
TI. ~ but only to have the relationship between derivatives:
(3.55)
Nonlinearity in Statics 93
Derivatives are calculated by the following formulae
(3.56)
The matrix belonging to an element e, e.g. the stiffness matrix, may be

expressed as
[h]e = r [B]e dxdy (3.57)

O'
where
B.. = aNi a~ + aNi a~

'J ax ax ay ay
aN]
ax' = [J] -; [aN]
CJ1l'
[-aNi
ay
aNi
- a~
M IINi
~
LJ ~ M aNi ]
~
- X i LJ -Yi
[J] = i=1 CJ1l i=1 CJ1l (3.58)
M iJN M iJN
[
~ ~j:'Xi ~ ~j:'Yi
II..
,=1 ,=1 II..
where [J] is the Jacobian matrix.

Taking into account the equations for isoparametric elements, formula
(3.57) is transformed to
I 1
[h]e = rr [B]eIJI dll d~ (3.59a)

-I -I
and for numerical calculation

NIP
[h]e = ~ WI [B/]eIJII (3.59b)
1=1
where [B/]e and 1JII are calculated at points and multiplied by weight coeffi-
cients Wit and NIP is a number of integration points.
3.5 NONLINEARITY IN STATICS [3.14; 3.32; 3.38; 3.43; 3.44]

In this Chapter FEM will be applied to solve nonlinear stationary problems
on the basis of:
• Newton-Raphson method, and

• direct iterative method.
Newton-Raphson's method
For nonlinear media, the stiffness matrix coefficients are functions of u. u
itself satisfies a system of nonlinear algebraic equations, Le.
[H(u)][u] - [8] = [0] (3.60)
The left-hand side of eqn (3.60) may be written as
[F(u)] = [H(u)] [u] - [8] (3.61)
Equation (3.61) is solved by Newton-Raphson's iterative method. After n
iterations the (n + l)th solution of eqn (3.60) is given by
[u n+ 1 ] = [un] + [L\u n] (3.62)
n I
The function [F(u + )], defined by eqn(3.61), may be expanded into a
Taylor's series:
[F(u n+ 1 )] = [F(u n)] + o~~~"l] In [L\u n] + o:[~~'l] In [L\~nF + ...

(3.63)
Truncating series (3.63) to two terms we obtain Newton-Raphson's method
[HT(u")] [L\u"] = [F(u n)]
(3.64)
[u n+ l ] = [un] + [L\u n]
where
[H ( n)] _ 0 [F(' )]
T U - o[u]
I n
(3.65)
is a Jacobian (tangential) matrix. The basic version of Newton-Raphson's

algorithm is shown in Fig. 3.9.
Direct iterative method

The direct iterative method is based on successively solving a system of linear
equations. Equation (3.60) is tranformed to
(3.66)
and solved by the algorithm shown in Fig. 3.10.
The Newton-Raphson's method and the direct iterative methods will be
explained by using examples of magnetic field calculation (thus u = A). The
Newton-Raphson's example uses a magnetic field calculation, excited in non-
linear media by direct current.
Let us assume, that
(3.67)
Starting
values
[u] = [ UO ]
Calculate
[HT (u·) ]]
Calculate
n
[F(u ) ]
Solve
[ HT ( U·) ] [ b. U·] = [ F ( u·) ]
[u· +'] = [u·] + [b.u·]
No
Fig. 3.9 Flow chart for Newton-Raphson's method
where:
k unit vector in the Z-direction,
A magnetic vector potential,
J current density vector.
The starting point for solving the assumed example is the first Maxwell's
equation
curlH =J (3.68)
curiA = B (3.69)
B = J!(B)H (3.70)
1
V=-
J!
Starting
values
[ u ] = [uO ]
Form
[ S ]
Calculate
[H(u')]
Solve
[H(u')][u"']=[S]
No
Fig. 3.10 Flow chart for direct iterative method
where v(B) is an explicit function of B or H (known as reluctivity), we obtain

a nonlinear Poisson's equation:
curl (v(B) curl A) = -J (3.71)
Taking into account eqn (3.67), eqn (3.71) is transformed to
div(v(') grad A) = -J (3.71a)
or
V' (v(·)VA) = -J
The boundary r, surrounding the domain Q, will be divided into two parts.
Two boundary conditions (Dirichlet or Neumann's) are applied in magne-
tostatics. For simplicity it is assumed that the boundary conditions are
homogeneous
A =0 (3.72)
iM =0 (3.73)
on
where n is the normal to the boundary r 2'
In Newton-Raphson's method the fundamental problem is to find the
Jacobian matrix (3.65). Thus
a[F] a a
[HT(u)] = a[u] = a[u] ([H(u)] [u]) = a[u] [H(u)] [u] + [H(u)]
= [T(u)] + [H(u)] (3.74)

because
a (3.75)
a[u] [H(u)] [u] = [T(u)]
The shape of matrix [n depends on how reluctivity is defined. In the case

when reluctivity is
v(·) = v(B) = v( IV x AI) (3.76)
the coefficients of matrix [n are given by
iij =J [VN;]T :; v'[(V[N][u]T)(V[N][u])]V~dO (3.77)
O' -
For the direct iteration method, coefficients of the stiffness matrix and load
matrix must be calculated
hij =J [VN;] TV (' )VNj dO (3.78)

O'
Sf = J JNidO (3.79)
O'
The next step is the reluctivity formulation. For this purpose the magnetiza-
tion curve is approximated by a multisegment method. The magnetization
curve is given by points, and successive points define the beginning and end
of successive segments. Reluctivity is constant inside intervals. Starting with
the first point, in defining the curve, we obtain
VI = V2 (3.80)
Hi
Vi =-,
B
i = 2.3, ... , N + 1
i
Reluctivity inside respective segments is calculated by the formula
v(B) = BVi+1 - Vi
(B - B i ) + Vi (3.81)
+ i 1 -Bi
where B belongs to [Bi• Bi+d.

B (T)
1,110
1,030
0,72
0,360
0
0 0,3 0,6 0,9 1.2 4
x10
H(A/m)
(a)
4
X 10
1.2714
~
E
-.
-;>
§:
~ 0,9981
0,7247
0,4514
0,1781
0 0,400 0,800 1,200 1.800
B(T)
(b)
2,5407
f 2,0325
...0
... 1,5244
;>I~
~~
1,0163
0,5081
0
B(T)
° 0,400 0,800
(c)
1,200 1.800
a A=O
v (.)
ad. - 0
iJy -
0} A=O
-a 10 a
..
x
iJA - 0
iJy -
-a
(d)
o a x
(e)
Fig. 3.11 (a) Magnetization curve, (b) reluctivity curve, (c) reluctivity curve gradient,
(d) infinitely long ferromagnetic conductor, (e) magnetic field distribution
The gradient of the reluctivity curve is defined by coefficients

V·+I - v·
a·=' , (3.82)
, B;+I -B;
with the following assumptions:
• if B = 0 then av = 0,
aB
• a;
avJ ; is the gradient for B;- = B·1+ I 2+ B·' ,
= [aB
• the gradient becomes linear inside successive segments defined by B,
• for B > Bsalural;on the gradient is equal to zero.
Thus for flux density B

B; ~ B ~ B;+I
the gradient of the reluctivity curve is given by the formula
av = a;+1
~B -
- a;
-
(B _ B-).
,+ a,. (3.83)
u B;+I -B;
In Fig. 3.11a-e the magnetization curve [3.11], reluctivity curve and reluctivity
curve gradient are presented for the steel containing 0.30/0 C. The method is
illustrated by a magnetic field calculation in an infinitely long ferromagnetic
conductor (Fig. 3.11d). As the conductor is symmetrical, the magnetic field
was calculated for a quarter of the cross-section (Fig.3.11e).
3.6 LINEAR TIME-DEPENDENT PROBLEMS

13.14; 3.32; 3.38; 3.40]
The FEM described in this Section is applied to time-varying electromagnetic
field. As there is a conducting media the problem can be described by a non-
homogeneous partial differential equation
aP
-- (au] aP
- -- (au] 2 au
- +k-=f (3.36a)
ax x ax ay Yay at
with the following boundary conditions
au
Pnan = UI(t) on f l (3.84)
u =0 on f 2
au = 0 on f 3
an
and the initial condition
u(l=O) = uo(x,y) for (x, y) belonging to n (3.85)
Linear Time-Dependent Problems 101
where
r1 + r2 + r3 =r
As in Section 3.5, a solution for time-varying problems will be presented
for the magnetic field calculation described by the equation
iM
V· (vVA) = a iii - J (3.86)
with boundary conditions

aA
v an = h(t) on rl (3.87)
A =0 on r2
aA = 0 on r3
an
and the initial condition
At=o = Ao(x,y), (x,y) belongs to Q (3.88)
The domain under consideration is presented in Fig. 3.12. Equations (3.86)-
(3.88) are substituted according to the weighted residual method, by the
following integral equation
!
o
WI V· (vVA) dO - ! aa:vt
0
WI dO + ! Jo + ! [v aAan - hJ
0
WI dO
rl
W2 dr =0
(3.89)
v· -
aA =h(t)
an
A=O
Fig. 3.12 Domain under investigation for time-dependent problem

After assuming, that

A = [N]T[u] (3.90)
wl=Ni
W2 = -ONi
applying the second Green's identity and Galerkin's method, eqn (3.89) is
transformed to
JVNlvv([N]T[u])dQ+ J ONi[[N]Td[u]]dQ_ J N;JodQ

o 0 dt 0
- J hN;dr = 0, i = 1,2,3, (3.91)

f,
or
[H][u] + [C] d~~] = [s] (3.92)
where
hij = J VNlvV~dQ (3.93)

O'
sr = J NiJodQ + J hNidr
O' f.
and [Cj is the conductivity (mass) matrix.

The boundary value problem (3.36a), (3.84), (3.85) will be solved by the
one-step 0 (theta) method. After the assumption
[un] = [u(nM)]
[u +0 ] = (I - 0)[u n] + 0[u n+l ]
n (3.94)
d[u] [u n+ l ] - [un]
- - ... for t = (n + 0)M
dt M
equation (3.92) for the instant t = (m + 0)l1t is written as
[L [C] + 0[H]J [u n+ l ] = [sn+0] + Ut [C] - (I - 0)[H]J [un]

(3.95)
with the initial condition (3.85). The 0 method is absolutely stable for
-! < 0 < I. For 0 = !, the accuracy is of second order (in time), while for
*"
o ! the order of accuracy is one.
The kind of 0 method to be used may be selected according to the values
of 0:
o =0 explicit Euler's method,

0=! Crank-Nicolson's (trapezoidal) method,
Linear Time-Dependent Problems 103
Starting values
[ UO ]
Set step length

t = 1i t
Calculate
[C']=l[C]
!it
Form
[S "tl]=[S"]_[H][u"]
Solve
[ c' ( 1iu") ] = [ S· ]
[ u • + 1 ] = [ u· ] + [ 1iu· ]
No
Fig. 3.13 Flow chart for explicit Euler's method
e =i Galerkin's implicit scheme,

e =1 implicit Euler's method.
Explicit Euler's method

The explicit Euler's method will now be used for the solution of the problem
described by eqns (3.86)-(3.88). In this case eqn (3.95) takes the form:
~t [C][u n+ = [8] +
1
] Ut [C] - [H]J [un] (3.96)
Equation (3.96) may be solved by a consistent mass matrix or by diagonal

mass matrix. The flow chart for the algorithm of consistent mass matrix is
presented in Fig. 3.13. The above algorithm is based on the following set of
equations
~ [C] [L\u n] = [sn] - [H] [un] (3.97)
[U n+ 1] = [un] + [L\u n] (3.98)

where
Cij = J aNi~dO
0'
hij = J VNrvV~dO (3.99)

0'
sr'" = J NrdO + J hnNdf

,
0' rl
In order to use the algorithm of the diagonal mass matrix, the matrix:
CII CI2 ]
[c] = ..
[ .... CiiCij ..
................
in eqn (3.95) is replaced by a diagonal matrix
d ll ]
[D] = d
22 : .. ,
[ ........d ll . .
.............
Matrix coefficients are calculated from relationships
T
dii=Ciia (3.100)
d..
'J
{*
= = 00 for i =j
for i *j
where
a=~Cii
i
For the majority of problems (except axisymmetric) T= 1. This mass matrix

formulation simplifies calculation because it does not require a solution of a
system of equations.
3.7 NONLINEAR TIME-DEPENDENT PROBLEMS

[3.7; 3.12; 3.38; 3.44]
As in Sections 3.5 and 3.6, magnetic field calculations will be used to explain
the solution of nonlinear time-dependent problems. For the domain, shown
in Fig. 3.14, the following space-time description is applied
Nonlinear Time-Dependent Problems 105
0
A=Aft
(!) Fe v aA
an Ir = h(t)
!-L,a I
o
air
Fig. 3.14 Nonlinear time-dependent problem (y-conductivity)
V' (v(')VA) = a~~ - Jo (3.101)
oA
v - = h(t) on f
on l
A = g(t) on f 2
A (x,y, t = 0) = Ao(x,y)
For any weight functions WI and W2 the following equation is satisfied
1{( V' (v ( .)VA» - a aa~ + Jo} WI dO + !

(v ( . ) ~~ - h (t)] W2 df =0
(3.102)
After applying the Green's first theorem
r V· (v(· )VA)wI dO = - r VAv(' )VWI dO + ; v(·) :A WI df

o 0 r un
(3.103)
eqn (3.102) becomes
j(} (-VAV(')VWI - a eMat WI + JOWl] dO + ~r v(·) ~Awldf

n
+
+ J(v(.)~ - h (t)J W2 df = 0
(3.104)
Assuming, that W2 = - WI and applying trial function WI which satisfies the

condition
we finally obtain
1 [VAV(')VW1 + (J ~~ WI - JOWl] dO - Jh(t)wl df =0 (3.105)
After assuming that WI = N;, eqn (3.105) transforms to
I [VN;V(' )V([N]T[u]) + (IN/([N~;[U]) - N;Jo] dO - JN;h(t) df = 0

I
(3.106)
for FEM eqn(3.103) is transformed to
(J [T(' )][u] + [H(' )][u] = [R] (3.107)
where
[ti] = d[u] (3.108)

dt
tij= J N;~dO (3.109)
0'
hij = J VN;v(· )V~dO (3.110)

0'
rj= J N;lodO + J N;h(t) df (3.111)

0' n
Equation (3.106) is solved by the E> method. In the calculation the following
relationships are used
T
tm = ml:1t, I:1t = - (3.112)
M
u'!' = u;(tm ) = u;(ml:1t) (3.113)
. [u m+l ] _ [um]
[u] == M (3.114)
[um+E>] = (1 - E>)[u m] + E>[u m+I ], (3.115)

Hence the set of equations (3.107) becomes
m1
(J [T] [u + ~~ [um] + [H(' )m+E>][um+E>] = [Rm+E>], m ~0
(3.116)
Nonlinear Time-Dependent Problems 107
The set of equations (3.116) is nonlinear. These equations will be solved by

predictor-corrector (P-C) method. In step P eqns (3.116) are written as
(3.117)
and in step C:
m l
a[T] [u + ] - [um] + [ir+I(')][um+0] = [Rm+0] (3.118)
M
where
[u m+0 ] = (1 - e)[u m] + e[u m+1 ] (3.119)
[u m+0 ] = (1 - e)[u m] + e [u m+l ] (3.120)
[H m(.)] = [H(u m)] (3.121)
[il m + 1 (.)] = [H(u m + 1 )] (3.122)
The problem formulated in this way requires the solution of two sets of
equations at the same time: one for extrapolation (P) and one for interpola-
tion (C). For e = 1/2, Crank-Nicolson's method is used. This method is based
on the Richardson's extrapolation [3.12]. Let E be the mean extrapolation
operator
m ~ 1 (3.123)
Taking into account eqn (3.115), eqn (3.116) is transformed to
m l
a[T] [u + ] - [um] + [H(Eu m)][u m+l12 ] = [R m+1/2] (3.124)
M
where
(3.125)
For this formulation eqn (3.124) is solved in one step only, but the solutions
in the two previous steps have to be known.
One of the best methods for solving eqn (3.116) is the Newton-Raphson's
method [3.33].
Newton-Raphson's method
This set of eqns (3.117) is modified before applying Newton-Raphson's
method. For this purpose the following formulation is used
m+0] [m]
[u m+l ] _ [um] = [u e- u , (3.126)
. [u m+0 ] _ [um]
[u] "" eM (3.127)
(3.128)
Taking into account eqn (3.127), the set of eqns (3.116) is transformed to
[ [H(' )m+0] + ~
~te
[T]J [u m+0 ] = ~
~te
[T] [um] + [R m+0 ] (3.129)
or for brevity
(3.130)
Denoting [F(·)] as
[F(')] = [K m+0(')][u m+0 ] _ [sm+0] (3.131)
and developing [F(·)] in k + 1 iterative steps into a Taylor's series
a[F(')]
[F(·)]lk+l= [F(·)]lk+ a[u] k[~Uk]+'" =0
I (3.132)
the Newton-Raphson's method is formulated as

[K1!+0(. )]d~Uk] = [sm+0] - [K m+0 (. )]k[Uk'+0] (3.133)
[Uk'':10] = [uk'+0] + [~Uk]
where
[K m + 0 (.)] = a[F(')] Im+0 (3.134)

T a[ul k
is the tangential matrix.
After solving in time (m + e)M, solution in time (m + l)M is calculated
from eqn (3.120). The tangential matrix is calculated in the following way
[K]!+0(.)] = a[u;+0] [[[H m+0(')lk + ~;e [T1J [uk'+01J (3.135)
For simplicity, indices m + e and k are ignored. In order to calculate coeffi-

cients k~ij of the matrix [KT (')] the following derivatives are calculated
aa.[fVN;V(')V([N]T[U])do+
uJ 0<
~ t a+ ef 0<
N;([N]T[u]) dO}
= f VN;v(')VNj[u]do+ f VN/~(')V([N]T[U])do
0< 0< uJ
+~fNNdo
~te J I
(3.136)
0<
At the start of the slope of the reluctivity curve av(' )/aUj is calculated. If the
reluctivity curve is given as a function of the flux density B the tangential
Boundary-Element Method 109
matrix [KT (')] is symmetric. This matrix is also symmetric if v = v(B 2). For
calculation it is easier to use v =v(B 2) than v =v(B):
B 2 -- Bx2 + By2 (3.137)
B = a([N]T[u]) . B = a([N]T[u]) (3.138)
x ay , Y ax
av(·) = av . aB
2
= dv (2a([N]T[U]) a1\j + 2a([N]T[u]) a1\j1
:2
2
aUj aB aUj dB 2 ay ay ax axJ
=2 V([N]T[u])V1\j (3.139)
After substituting eqn (3.139) into eqn (3.136) we obtain
ke.....
.IJ
= Jr VNv(')VNdO
JI Me Jr NNdO
+~ J I
o O'
+2
O'
I VNjV([N]T[U]) ~2 V([N]T[u])V1\jdO (3.140)
In Fig. 3.15 the flow chart for the predictor-corrector method is presented.
3.8 BOUNDARY-ELEMENT METHOD

(3.4; 3.5; 3.6; 3.24; 3.25; 3.35]
The bases for the BEM formulation are Green's second theorem:
o
I (uaG - Gau) dO = ;r (u aGan - G~uJ
un
df (3.141)
and the Green's function G. For the Laplace operation a in two-dimensional

space:
1 1
G = -In- (3.142)
21t r
and in three-dimensions:
1
G=- (3.143)
41tr
where:
r distance between the source and point under consideration.
Let us derive the boundary-integral equations. Figure 3.16 shows space 0
surrounded by a boundary surface f. Inside the domain 0, Poisson's equation
V2u = - f is satisfied. Let us assume that on f. Dirichlet's conditions hold
u = Uo and on f 2 Neumann's conditions apply au/an = <Po. The Green's
function satisfies the following equation
DATA:
e, !!J, uo
m=O,i=O
(]
[KJ=Ai[TJ+8[H(»
(]
[P J =-[ T J. (l - 8)[ H ( .
lit
»
Yes
No
Fig. 3.15 Flow chart for predictor-corrector method
(3.144)
where
o(P) = Dirac function
'* 0 for r = p
o(P) = {= 0 '*
for r p.
Formula (3.141) takes the form
u{P} = Jr GjdQ - 1l. u-;-df

oG + 1l. G-;-df
au (3.145)
o r un r un
which allows calculation of u(P) inside domain Q on the basis of u and au/an
on the boundary f.
Boundary-Element Method 111
Fig. 3.16 Introducing BEM
Equation (3.145) will now be derived. For this, circle K with radius p is
removed from domain O. In consequence of this for eqn (3.141) an integral
along the boundary f k integration must be added
J
C)-x
{u~G - G~u} dO = f [u-aGan - G-anaUJ df + f (aG
r an
aUJ df
u- - G-
an r.
k
(3.146)
After taking into account that in the region 0 - K:
V 2G =0 (3.147)
we obtain:
J
O-K GjdO =1
L (aG
u an - G an df
aUJ + ~L [aG
u an - aUJ
G an dfk (3.148)
Now an integral along the boundary f k' when p -+ 0, will be calculated:
lim
p.... O
I u{p) :G df = lim J u{p} -21 (- ~JP pdq> = - u{P)
lin
k
p.... O
2lt
1t
r. 0
where
and
~u Ip G dfk = lim J ~u I ~ In!P p dq> = 0

271
lim J lin
p-+O fk p-+O 0 lin p 21t
because
· p In -1
I1m
p-+O p
= - p-+O
I'1m p In p = - I'1m-
In p
p-+ol/p
and after application of d'Hospital's principle
· -In p
- I1m = lip-+O
m -1/ p I'
- 2 = 1m p =
0
p-+O 1/P 1/ p p-+O
In the surroundings of point B (Fig. 3.16) integration follows the semicircle

(q> belongs to (0,1t», then
lim J u(B) ~G dfk = -~U(B) (3.149)

p-+O fk n
and
Thus eqn (3.146) reduces to eqn (3.145) inside 0, whereas on the boundary
1
-u(B) - J GfdO = 1t [ao
u- - aUJ
G- df l (3.150)
2 a f an an
Finally, by applying boundary conditions, we find
The boundary-element method is a numerical procedure for solving boundary-

integral equations. For this, boundary f is divided into elements: f l into N
and f 2 into M elements. Various approximations may be used on these
elements. We will obtain the simplest formulae when it is assumed that the
function u or its derivatives are constant along the element.
Then (Fig. 3.17)
(3.152)
(3.153)
Finally formula (3.151) transforms to

Open-Boundary Problems 113
f,
Fig. 3.17 Boundary discretization
where i is any point on the boundary f.

For N + M unknown values of function or their normal derivatives,
eqns (3.154) are formulated. For calculation of coefficients occurring in
eqn(3.154), Green's function (3.152) or its derivatives (3.153) and three
integrals occurring on the right-hand side of eqn(3.154) must be calculated.
The given quantities are
3.9 OPEN-BOUNDARY PROBLEMS 13.2; 3.3; 3.27; 3.38)

Sometimes it is necessary to calculate the electromagnetic field in an
unbounded region (Fig. 3.18). In a bounded region (1) Laplace's, Poisson's,
Helmholtz's or Fourier's equations may be satisfied. Media existing in region
(1) may be linear or nonlinear. All the sources are in the bounded region. In
an unbounded (2) region Laplace's equation is satisfied.
Let us solve the problem shown in Fig. 3.18 by interfacing FEM and BEM.
The bounded region (1) is divided into finite elements, whereas in the exterior
region (2) BEM is used. On the interface f the boundary conditions are
unknown. For region (1) FE equations are formulated and BE (nodes
i = 1,2, ... , M) are added to them. Two sets of equations (FE and BE) are
combined together and solved on a computer.
For elements which do not have nodes on the border f, the stiffness matrix
[H] is calculated in the way shown above. Numbering of border elements is
clockwise. For this kind of numbering the normal unit vector (0) is directed
from region (2) to (1). Let us discuss elements placed on the border f. A
~u=o
Fig. 3.18 Electromagnetic field calculation in unbounded region
segment connecting nodes i, i + 1 placed on the border r is the ith boundary

element r j (Fig. 3.19).
First, let us assume that Neumann's boundary condition is known on the
boundary. For this element
(3.155)
Coefficients of the vector [pel are calculated from
Ff = f Njqdrj (3.156)
fi
Ff+1 = f Nj+1qdrj (3.157)

fi
where:
q normal derivative,
Nj,Nj + 1 shape functions on the boundary r j •
= -! 0 - ~)
Nj (3.158)
N + 1 = to +~)
j (3.158a)
L
~ belongs to (- 1, 1), dr = ~d~
2
where L j is the element length.
The minus sign in eqn (3.155) results from the opposing direction of the unit
normal vector. After assuming a linear approximation of the derivative on the
border
(3.159)
we obtain
Fi = J N; df;q; + J N;N;+ 1 df;q;+ ( (3.160)

ri rj
Fi+1 = J N;Ni+1 df;q; + J N;+l df;q;+1 (3.161)

ri ri
After integration
(3.162)
Global matrix coefficients are calculated from the formula
(3.163)
where
The global matrix is defined by a formula

[F] = [A] [Q] (3.164)
Matrix [A] has a dimension M x M and reads
L 1 +LM L( LM
0 0 0
3 6 6
L( L( +L2 L2
0 0 0
6 3 6
[A] = L2 L2 + L3 L3
0 0 0
6 3 6
o o o
(3.165)
Vector [Q] has dimension M and contains normal derivatives. The following
boundary integral equation can be written in the region (2)
M 00 M 00]
+ ~ J u-;-dfj = ~ J qOdfj + lim J a
2n [
Cju; qO ~u Rd<p
j=1 rj un j=( rj R-oo 0 n (3.166)
where:
1 1
0= -In-
21t r
any point on the border f,

j boundary element number.
The last integral of the right-hand side of eqn (3.166) is equal to zero. A similar
situation exists for the application of the BEM to bounded regions. Coeffi-
cients C; do not have to be calculated explicitly. However, there is a differ-
ence in the way that the elements of the matrix [H] are determined. For this
purpose we will assume that potential u is constant in domain (2), hence the
right-hand side of eqn (3.166) will not be equal to zero as in the interior
domain. This follows from
21t 21t
aG dq>
· Jr u-R
11m = l'1m u- -1 Jr-R
i d q> = -u (3.167)
R.... oo 0 an R-+oo 27t 0 R
Thus
M
H;;=-~ Hij+ 1 (3.168)
)=1
)*;
where i = 1,2, ... , M.

The determination of Hij and Gij is the same as in the classical version of
the BEM. Vector [F] can be determined by the boundary r potential [url:
[F] = [A][G]-I[H][ur] (3.169)
It will be noticed that the product of [A], [Gr l and [K] is a stiffness matrix
for the macroelement with Mnodes placed on the boundary r (Fig. 3.19). This
matrix is asymmetrical.
In recent years a number of special elements have been introduced in order
to extend the scope of the standard finite-element method. 'Infinite finite-
elements' have become increasingly used and accepted as a means of solving
unbounded field problems [3.18; 3.19; 3.27]. The interior region (1) is
modelled by finite elements and the exterior region (2) is represented by
infinite elements. A very simple infinite-element for a two-dimensional prob-
lem was proposed by Pissanetzky [3.27]. This element has two nodes and con-
forms with linear triangles and linear quadrilaterals.
In Fig. 3.20a subspace 0; is divided into finite elements and in Fig.3.20b
the infinite element is shown. This element has two points 1 and 2 defined
by coordinates (XI' Y 1) and (X2 , Y2). The auxiliary coordinates system (v, w)
is introduced. Axis v is parallel to segment 1-2. For calculation, the following
functions are used
/ = -!-v'[(X2 - X I )2 + (Y2 _ YI)2
1
d = 2/ (XI Y2 - X 2 Y I)
Fig.3.19 FEM and HEM connection
b = .!
41
(xl - x/ + yl - y/) (3.170)
rm = .J(d2 + b 2 )
1
tan<P1 = d (b -I),
Additional axis II is introduced by the formula:
1
II =I (dtan<p - b) (3.171)
where
(3.17Ia)
The function u is linear between points 1 and 2:

UI-2(1l) =! [(I -ll)Ul + (I + 11)U2]. (3.172)
In terms of the polar coordinates, points belonging to segment 1-2 have
coordinates (d/cos <p, <p). If one wants U to decay as a given function f(r)
towards infinity for any <p = constant, it is enough to assume in the infinite
element
fer)
U( r, <p) = 2f( d/cos <p) [( 1 - ll) U1 + (I + 11) U2]
Thus the shape functions for the infinite elements are
(a)
(b) y
Fig. 3.20 (a) Division of domain into finite and infinite elements, (b) two-nodal infinite
element
f(r)
N. (r, q» = 2f( dl cos q» (1 - TI)
(3.173)
f(x)
N 2 (r,q» = 2f(dlcosq» (1 + TI)
Let us consider the particular case when f(r) = l/r n , where n ~ 1. Taking
into account eqns (3.171) and (3.173) and the relationship
tan q> = (rsin q»1 (rcos q» = wi v
we obtain
N 1 (v, w) = ~n (1 + ~ - dWJ
2v / /v
(3.174)
N 2 (v, w) = ~n (1 - ~ - dWJ
2v / /v
For solution of the above mentioned problem, the global functional has to
be minimized in subspaces (1) + (2). In subspace (2) Laplace's equation is
satisfied:
(3.175)
The functional for Laplace's equation is the particular case for functional
(3.38) and is expressed by the formula
(3.176)
We will write this functional in the v Ow coordinates system. The relationship

between v Ow and x Oy systems is the following
rLYx] = [ - c~s
sm a
a sin a]
cos a w
[v] (3.177)
where a is the angle between coordinates v Ow and x Oy, and the inverse
relation is as follows
[v]
w
= [c~s a
sma
- sin a]
cos a
rLYx] (3.178)
Derivatives are expressed by relationships

au au ov au ow
-=--+--
ox ovox ow ox
(3.179)
au au ov au ow
-=--+--
oy ovoy away
or in a matrix form, after taking into account eqn (3.178)
::] [cos a sin a] [:~] (3.180)

[ :; = -sina cosa : :
From eqn (1.179) it follows, that
(3.181)
Because
J f(x,y) dxdy = J g(v, w) IJI dvdw (3.182)

o 0
where f(x, y) and g(v, w) are alternative expressions for functions in coordi-
nates xOy or vOx and IJI is the Jacobian matrix,
dxdy = IJI dvdw = 1 a(x,y) I dvdw = I ax/av ax/aw I dvdw

a(v, w) ay/av ay/aw
=1 -sma
c~sa sinaldvdw=dvdw
cosa
(3.183)
and finally the functional (3.176) can be written as
(3.184)
For nodes 1 and 2, after differentiation of the functional (3.184) we obtain
aet>e = 2J [au !.... [au] + au !.... (au]] dvdw, j = 1,2 (3.185)

au; e
av au; av aw au; aw
where
au = [aNI, aN2] [u 1]; (3.186)

av av av U2
!... (au] _ aN; (3.187)

au; av - av '
Taking into account the integration limits for variables v and w
W{(b + l)v/d (3.188)

(b -l)v/d
for the whole IE we obtain
Finally, the stiffness matrix is found as
[H] e = 6~d [.~.4::·:·~~~~(l.~·~:t~;~:~:~~/±·;;;] (3.190)

sym ! + 6nb + 3rVI
The closed-form expressions for the element matrix can also be obtained for
the infinite element which has three nodes and conforms with 'quadratic
quadrilaterals' [3.18].
Figure 3.21 shows an example internal subdivision into finite elements and
external subdivision into infinite elements.
The geometrical relationships for 3-node infinite elements are the same as
those for 2-node elements. In the middle of segment 1-2 a third node (marked
as 3) is added with appropriate description. Other notation is the same as in
Fig. 3.20. The unknown function u has values u .. U2' U3 in nodes 1, 2, 3,
respectively and varies parabolically along the segment 1-2-3
v
Fig. 3.21 Internal subdivision into finite elements and external subdivision into infinite
elements
(3.191)
The shape functions for the three nodal infinite elements, in terms of their
polar coordinates r, q> are
-j(r) 1 -11
N. (r, q» = j(d/cos q» 11-2-
j(r) 1 + 11
N 2 (r,q» = j(d/cosq» 11-2- (3.192)
j(r) 2
N 3 (r,q» = j(d/cosq» (1 -11 )
Coordinate 1/ is given by formula (3.171). Let us assume that as for 2-nodal

infinite elements
1
j( r) = Ii'
r
n ~ 1.
We will write the shape functions in local coordinates
N 1 (v, w) = ~ [~ - ~I ~Jv [1 - ~I ~v + ~JI

2v n I
N 2 (v, w) =~ (~ ~ - ~JI [1 + ~I ~v - ~JI

2v n I v
(3.193)
N 3 (v, w) = d n [1
v
n
_r~LI ~v - ~J2]
I
Thus, after calculation, we obtain stiffness matrix elements for 3-nodal
infinite elements
v
(b+J) d 00
H IJ~. = 2 J J (aNi
av a~
av + aNi
aw aN
J
aw ] d v dw , i = 1, 2, 3
; '
J = 1, 2, 3
(b-J)£ d
(3.194)
After a simple but very laborious calculation the following result is obtained
H11 = -30n1d [(8n 2 + 22n + 17)/- IOb(3n + 4) + 35r~/I]

1
Hh =- d [( -2n 2 + 2n + 7)1 + 5r~/I]
30n
H13 = -15n2d [(n 2 - n - 6)1 + lOb - IOr~/I],
H~I = Hh, (3.195)

1
H~2 =- - [(8n 2 + 22n + 17)1 + IOb(3n + 4) + 35rVI]
30nd
Hierarchical Elements 123
H~3 = 15~d [(n 2 - n - 6)1 - lOb - IOr~/l]

Hil = H13
Hi2 = H e23
8
Hi3 = 15nd [(2n 2 - 2n + 3)1 + 5r~/ l]
3.10 HIERARCHICAL ELEMENTS [3.1; 3.23; 3.38; 3.39]

The hierarchical elements are used in computation of electromagnetic fields
by the finite-element and bound~ry-element methods. They are very useful in
local improvement of the accuracy of calculations without changing the finite
elements mesh. Local improvement can be achieved by:
(1) local high discretization of the finite elements mesh,

(2) application of hierarchical elements.
Let us discuss hierarchical elements. In this kind of element shape functions

are applied in which increasing the order of the polynomial does not break
the continuity of the interface between elements. Accuracy of calculations
increases by increasing the order of the polynomial without increasing mesh
density. One polynomial for this type of application is the Hermitian poly-
nomial. These are used in the solution of the following differential equation
[3.20]
(3.196)
where n = 0, 1, 2, ....
H,,(x) = (- I)n ex2 :;n (Xe-xl)
= 2nx n - 2 n- 1 [~J x n- 2 + 2 n- 2. 1 ·3· [:J x n- 4 - ••• (3.197)
For n = 0, 1,2,3,4 and 5 Hermitian polynomials have the following forms

Ho(x) = I; HI (x) = 2x; H 2(x) = 4x 2 - 2
H 3(x) = 8x 3 - 12x; H 4 (x) = 16x 4 - 48x 2 + 12
Hs(x) = 32x s - 160x 3 + 120
Hermitian polynomials satisfy the following recursive relationships
Hn-1(x) = 2xH(x) - 2nHn_ l (x)
dHn(x)
dx = 2nHn_ 1 (x) (3.198)
Fig. 3.22 The discretization of a model of a transformer leakage area into elements
In reference [3.23] Hermitian polynomials are used for a transformer leakage

magnetic field calculation. In Fig. 3.22 one quarter of a transformer cross-
section with finite elements is shown.
The vector potential is used for calculations. In the model, potential A has
only one component
A=A(x,y) (3.199)
satisfies the following equation

a2A a2A
ax 2 + ay 2 = joroJ,1.A - J.LJ in coils
a2A a2A
ax2 + ay 2 = joroJ,1.A in conductors (3.200)
a2A a2A in dielectrics

ax2 + ay 2 = 0
where J is the current density in the coil, and 0 is the conductivity.

Vector potential A was interpolated by the following expression:
aA aA a2A a2A a2A
A (ll,~) = HooA + H IO all + HOI ~ + H II OTta~ + H 2O all2 + H 02 a~2
(3.201)
where:
Three-Dimensional Problems 125
(a) (b)
Fig. 3.23 Field lines distribution for (a) zero, (b) first and second order Hermitian
polynomials
T1, ~ local coordinates,

Hm,n Hermitian polynomials,
m,n 0,1,2.
In Figs 3.23a, b distributions of vector potential A are shown which are
obtained from FEM calculations using zero, first- and second-order Hermitian
elements.
3.11 THREE-DIMENSIONAL PROBLEMS

[3.4; 3.10; 3.19; 3.25; 3.38]
A more realistic image of electromagnetic phenomena is obtained using three-
dimensional analysis. In a number of practical applications a two-dimensional
approximation may be sufficient without causing significant error in results.
An application of FEM to two-dimensional calculations was presented in
Section 3.3. In this Section FEM is applied to three-dimensional calculations.
Thus in eqn (3.38) describing a two-dimensional scalar field, we will add a
derivative with respect to the third coordinate and the equation becomes
a[au] - -a[au]
-- P -
~ X~
p - - -a[au]
~
p - + k u =f
h zh y~
2 (3.202)
The functional, equivalent to the non-homogeneous Neumann's boundary

value problem for eqn (3.202) has the following form
~(u) = 1 [::r- [:;r- [:ir

[Px Py pz
22
+ k u - 2f U] dxdydz
-2! (j)(B)udf (3.203)

r
Fig. 3.24 Tetrahedral element
(3.204)
In Fig. 3.24 the simplest three-dimensional tetrahedral element is presented.

For this element m = i,j, k, I and
u'(x,y,z) ~ [N,(x,y,z), ~(x,y,z), Nk(x,y,z), N,(x,y,z)) [~]

(3.205)
Equation (3.29a) for the tetrahedral element has the following form
n
a4>e
au;
a4>e
-
aUj =
[h:J'
hij h1k
he. ht hjk h"l
J
hlU'j +u·J S.J (3.206)
a4>e h~; h~j h~k h~, Uk Sk
-
aUk hf; hlj h7k hfJ U, S,
a4>e
-
au,
Instead of eqn (3.28) we have

(3.207)
where m = i,j, k, I, and instead of eqn (3.30):
1
«\ = 6V (ajuj + ajuj + akuk + a,u,)
1
«2 = 6V (bjuj + bjuj + bkuk + b,u,)
(3.208)
1
«3 = 6V (CjUj + CjUj + CkUk + c,u,)
1
«4 = 6V (djuj + djuj + dkuk + d,u,)
where
6V = det [~ ~: ::
1 Xk Yk Zk
;:] (3.209)
1 X, Y, z,
and V is the volume of tetrahedral element i,j, k, I.
aj = det [~ ;~] bj = det [~ ~]

;,
Xk Yk Zk 1 Yk Zk
(3.210)
Cj = det [~ ~ ~] dj = det [~ ~~ ~]
Xk 1 Zk Xk Yk 1
The other coefficients are defined by right-hand cyclic permutation of the
subscripts in the order k, i,j, I.
Formula (3.36) will transform into
u'(x,y,z) = IN"M,N.,Nt] [~] = [N] T[U]' (3.211)
where
(3.212)
Formulae (3.41), (3.43), (3.44) and (3.45) will be transformed similarly. In

the same way, as shown in Fig. 3.5, the full equation for the three-dimensional
case is assembled. More complicated elements are also used.
p o
I
I
I
I
I
:I k
,..0--
"
"
Fig. 3.25 Linear hexahedral element
Fig. 3.26 Ten-nodal tetrahedral element
Figure 3.25 shows a linear hexahedral and Fig. 3.26 demonstrates a IO-node
tetrahedral element. For the linear hexahedral element eqn (3.28) becomes:
u
e
= Q1 + Q2 X + Q3Y + Q4Z + QsXY + Q6YZ + Q7 XZ + Qgxyz (3.213)
The calculations are similar to those which were done for tetrahedral elements.
In three-dimensional calculations isoparametric elements are applied in the
same way as calculations which were done in Section 3.4.
The simple example [3.17] of scalar magnetic field calculation will now be
presented. The model is shown in Fig. 3.27. Inside the ferromagnetic material
(J.1 :> J.10) there is an empty hexahedron (J.1 = J.10). On the surface between
regions with J.1 :> J.10 and J.1 = J.10 electric current flows with linear density K o•
We will calculate the magnetic field inside the air region. In this region the
following equations are satisfied
curlH = 0
divB = 0 (3.214)
8 =l-'oH
Magnetic scalar potential is determined by the relationships
H = -gradq> (3.215)
y
K.
/: /
b
,
-7 . c
/ / I
I
I ,
,
,
.~-- ... ~-
, ,,'- 'IJ • x
11 • 11 0 , , , I
I
,,
/' c
11 » 110
h
z
Fig. 3.27 Model for scalar magnetic field calculation
and fulfils Laplace's equation
°
02q> 02q> 02q>
(3.216)
ox 2 + oy2 + OZ2 =
Taking symmetry of the model into account, calculations were done for half
of the model presented in Fig. 3.27.
The following boundary conditions are prescribed:
n .B = ° for f x = 0, y e (0, b), z e (0, h)

LY = 0, X e (0, a), Z e (0, h) (3.217)
Hxn=K (3.218)
on the rest of the walls
Kof x=a,ye(O,b), ze(O,c)
=
K=
°
-Kol =
x
y
a,ye (0, b), ze (c, h)
b, x e (0, a), z e (0, c)
° == y b,xe (O,a), ze (c,h)
° z O,xe (0, a), ye (0, b)
° z = h,xe (0, a), ye (0, b)

Taking into account eqn (3.215) boundary conditions (3.217) and (3.218) are
expressed in terms of scalar potential. Thus condition (3.217) reads
oq>=° f = for x 0, y e (0, b), z e (0, h)

(3.217a)
on LY = 0, X e (0, a) , Z e (0, h)
and eqn (3.217) becomes
0.300 , , : , : , : : : :
. .. .
....... ~ ~ ; ~ ; ~ ; : : ; --'0--' H,
---A-_. H,
---{3-_. H,
---)(-_. H
Fig. 3.28 Numerical solution
q>=o for z = 0, xe (0, a),ye (0, b)

<p = -Koz for x = a,ye (O,b),ze (O,c)
<p = -Koc for x = a,ye (O,b),ze (c,h) (3.218a)
<p = -Koz for y = b,xe (O,a),ze (O,c)
<p = -Koc for y = b,xe (0, a), ze (c, h)
The results of the numerical calculations are presented in Fig. 3.28. The
problem has an analytical solution. For comparison, in Fig. 3.29, results of
the analytical solution are also presented. More information on the example
may be found in reference [3.17].
3.12 NUMERICAL SOLUTION OF LARGE SYSTEMS

OF EQUATIONS [3.4; 3.21; 3.37; 3.40]
Electromagnetic field calculations by finite element and boundary element
methods require the solution of large systems of algebraic equations. The
methods for solving such large systems of equations may be divided into two
basic groups:
• direct methods, and

• iterative methods.
Numerical Solution of Large Systems of Equations 131
0.300 ·· . . .
. .
.......•...........................................................
. .. .. .. .._ .. ...
....... L ; ~ ~ ~ ,- ~ ; ; ; : ---(3--- H.
: : : : : : : :
::: .~r\~;i:·F~j·.· • •I.·••··.k;j·...·~:·~r• •

:~
--06--- H,
--08--- H,
---)(--- H
1
· " 1Jt· " . '. . ~., . .
0.210
..... ! .;
0.180
0.150
0.120
0.090 A
........~ / .. ~ ~ j ~ ,.l [ \. \ ~ o.o.~
0.060 ~ .,' ; ~ ; {- ~~" ; ~ ; \ : ;
0.030 :::::)~:::::[:::::::;::::::::[:;/:r
is· . . .,
. \::I:::::::!:::::::I:::\1::::::::1
...
·~·~o t&.:; ~ _..~ IE;: ~0< _ :.~'fA: ~./};~.~ .. ~'}"'~ _.. ·LA.:~
0·
0.000
0.0 0.6 1.3 2.9 2.6 3.2
Fig. 3.29 Analytical solution
Direct methods
Let us consider an algebraic system of equations:
[k] [a] = [r] (3.219)
where:
[k] square matrix of coefficients,
[a] vector of unknown values,
[r] vector of right-hand sides.
Let the coefficient matrix [k] not require changing rows or columns. This
will be the case if [k] is symmetric and positive (or negative) definite. It may
or may not be true when a matrix is asymmetric or not definite. If row
or column changes are necessary, then the following procedure has to be
modified.
Let us additionally assume that [k] can be written as a product of two
matrices
[k] = [L][ U] (3.220)
where:
[L] a lower triangular matrix with a unity main diagonal,
[ U] an upper triangular matrix,
]
[L]
[I 0
= L~I ~
----
---- 0
L n1 L n2 1
(3.221)
U l2
[ V" U22 U2n
[vl= ~ - ---- V'"J
0 Unn
The above process is known as triangular decomposition of the matrix [k].
Thus the solution of eqn (3.219) may be expressed in terms of the solution of
two equations:
[L][y] = [r] (3.222)
[U][o] = [y]
The solution is very easy in this case and can be written as
YI = rl
i-I
YI = ri - L; LijY) i = 2,3, ... , n (3.223a)

)=1
on = y/Unn
0i t
= [Yi - }=i+1 UijOu] / Uii (3.223b)
The procedures described by eqns (3.223a) and (3.223b) may be referred to

as forward elimination and backward substitution, respectively. The first three
steps of the decomposition process are shown in Table 3.1.
Table 3.1 Triangular division of matrix [k)
Step 2:
[
k
::
ll k121k13[1 0 [Ull Ul2 =k l2 1
~~~_~;:J ~:~~_~:~~~~_~::_~! ~ C:::_~_~:2_~_~::C::: I
Step 3:
___JI kkl31
23 :
[1L 21 I
01 I 0 'r, 0Ul1 U22
U
l2 l3
U 23 =k
l3
= k 23 - L 21 Ul3
1 -------------- I
[ k 31k 32 k 33 1 L 31 = k 31 1Ul1 L 33 =1 I 0 0 U 33
----------1 L 32 = (k32 - L 31 UdlU22 I = k 33 - L 31 U l3 - L 32 U 23
Numerical Solution of Large Systems of Equations 133
The algorithm of decomposition of a square matrix of order n x n is as

follows
L l1 = 1 (3.223)
then for each active area (from 2 to n)
L jl = kjl Ul1 (3.224)
UIj = klj
followed by
(3.225)
i-I
Uij = kij - ~ LimUmi (3.226)

m=1
and finally
(3.225a)
j-I
Ujj = k jj - ~ LjmUmj (3.226a)

m=1
The decomposition process is shown in Fig. 3.30. After the single matrix
decomposition eqn (3.185) can be solved repeatedly for several right-hand side
vector [r] values. For a very large matrix [k] the process of decomposition may
be very time-consuming; thus repeated calculations without a need to perform
the decomposition is a valuable property. The above analysis was done for
general types of matrix [k]. In the case of FEM, coefficients of matrices have
special properties. Very often the matrix is symmetric (kij = k j ;), thus the
following relation is true
U
j·th column
active area
•
,,
I ,,
- .
,, leli ,, Uu Uli
,,
leZi ,, UZi UZi
'reduced ,,
L
area. , ,,
,, ,,
, ,, ,,
,,
,,
,, ,,
,,
--=- c...:...
,, ,
@J5J
j·th row
active leil lei' •.......k"
,,
LiI ,Lj2 .... ·1
area ,,
, I I
",:
unreduced
area
,,
,,
LwLp .... ·1 GI····,
Ie.J' ~',:
~:
__ .... _~ ..
0
,,
,
,
Fig. 3.30 Matrix decomposition

1/2 band 1/2 band
X XX X
XIXXX~
IX X X profile X XX
XXX X XX
XXX X XX
xx ~ X X
XXX X XX
XX X X
SYM
XXX X XX
~~
X
A
X
~
Fig. 3.31 Typical storage scheme for a profile matrix
Uij = Lj;U;;
Hence, for such problems, it is not necessary to store both matrices [L] and
[V] but only one of them (upper or lower triangular matrix). This causes
reduction by half of the required computer memory. We obtain the further
savings of memory if only non-zero coefficients are stored. In FEM the
maximum band width is usually equal to 10-20% of the number of unknowns.
Figure 3.31 shows a typical storage scheme for a profile matrix. Further sim-
plification is demonstrated by Fig. 3.32. More details about efficient storage
systems may be found in Section 7.4 of Chapter 7.
Iterative methods [3.37]

The two simplest methods are those of Gauss-Seidel and over-relaxation:
in Gauss-Seidel's method, in the first step, matrix [k] is divided into two
triangular matrices [L] and [U]
[k] = [L] + [U] (3.227)
where [L] is the lower triangular matrix defined by
-k .. {i=I,2, ... ,n
L·-
'J IJ j=I,2, ...,i
k
XIX X X ~ 1
IX X X ~ 2
XXX ~ 3
XXX ~ .-
XX ~ 5
XX
~
6
XX 7
SYM
XX X
~~
X
Fig. 3.32 Reducing computer memory

Applications 135
[U] is the upper triangular matrix defined by:

Vi} = ki} i = 1,2, ... , n - I
j = i + I, i + 2, ... , n
All the other elements of matrices [LJ and [U] are equal to zero.
The basic algorithm of Gauss-Seidel's method is described by the following
equations
[aO] = [v]
[L][a n+1 ] = [r] - [V][a n] (3.228)
where [v] is the initial vector and indices correspond to successive iterations.
If matrix [kJis symmetrical and positive definite then the Gauss-Seidel's
method is convergent, even though the process may be very slow.
In order to improve the rate of convergence over-relaxation may be applied.
Subtracting [L][anJ from both sides of eqn (3.228) yields
[L] [~a] = [r] - [k] [an] (3.229)
and
(3.230)
13 is the over-relaxation parameter. The value of 13 is between 0 and 2 (stability
limit) and the optimum value is problem-dependent. For 13 = I, the scheme
reduces to the classical Gauss-Seidel's method.
The main advantage of using iterative methods lies in the savings of com-
puter memory as a result of eliminating the process of matrix decomposition.
The main drawback is the unknown number of necessary iteratives to achieve
prescribed accuracy and a related problem of difficulties in estimating the
optimal value of the relaxation factor. The method becomes unreliable for
asymmetrical matrices.
3.13 APPLICATIONS [3.41; 3.42]

3.13.1 l?lectror.nagnet
A two-dimensional magnetic field analysis is conducted for an alternating
current electromagnet, shown in Fig. 3.33. The armature is assumed to have
a fixed position and the coil is supplied by injecting the following current:
i o = (I - e-t/To)/o (3.231)
where:
To time constant,
/0 steady current.
Super-elements have been generated manually (Fig. 3.33), followed by an
automatic mesh generator (Fig. 3.34). Owing to symmetry, only half of the
system has been modelled.
D c
III
CD
III
10 <I'
CD N N
N
10 15
67
A B
Fig. 3.33 Electromagnet cross-section and super-e/ements
The problem is described in terms of the vector potential A = A,k = Ak.

There are four regions with different material properties: I - air, 2 - armature,
3 - core, 4 - coil. The following equations hold
a2A a2A
ax2 + ay 2 = 0; in region I (3.232)
-a [v(·)-
aAJ a [v(·)-
+- aAJ =0-;
aA in regions 2 and 3 (3.233)
ax ax ay ay at
a2A a2A
ax2 + ay 2 = - iLoio, in region 4 (3.234)
where 0 is the conductivity, with the boundary conditions

A =0 (3.235)
on the symmetry surface r DA.
On the external boundary r ABeD infinite elements have been applied (see
Section 3.9). The following initial condition is assumed.
A(I=o) =0 (3.236)
Applications 137
Fig. 3.34 High density mesh
In Fig. 3.35 the B-H curve for steel E41 used in armature and core con-
struction is given.
Figures 3.36 and 3.37 show field plots under steady state and transient
conditions respectively.
3.13.2 Leakage field in a transformer

Taking symmetry into account one quarter of a transformer cross section is
considered (Fig. 3.38). We assume that:
• the core is made of nonconducting ideal ferromagnetic material,

• the screen is made of copper,
• screened parts of the tank and core have the same properties,
• the beam and non-screened tank parts are constructed from the same
steel.
Let us assume that the magnetic vector potential has one component only
2.00 ,,
,
,,
,
,
,, ,, , ,,
___________ 1". • • _ _ __ I". L .L ,
1 .50 ,, I
I
I
I
,
,
I
,
, I I I
I I I I
I I I I
, , I ,
I I t I
I I , I
, I t I
I I I ,
, I I I
f-- ,
I
I
I
I
I
I
,
-- - ----- - -~- -- -- - -- -- -~ ----- -- -- --~ - --- -- -----~-- - - --- -- --:

I I I I I
l.-J 1 .00 , , I I ,
m
I I I I ,
I I I I I
I I I I ,
I I I I ,
I I I I ,
I I I I ,
I , I I ,
I , I I ,
I , I I I
I I I I I
I I I I I
I , I I I
0.50 -----------~-----------~-----------~-----------~-----------:
I I I I I
I I I , I
I I I I ,
I , I I I
I I I I ,
I I I , I
I I I I I
I I , I ,
I I I I I
I I , I ,
I I I I ,
I I , I ,
o. 00 +rrrrrrrrr+TTTTTTTTT"tTTTTTTTTTt"TTTTTTT-rr+TTTTTT""TTT""1
I I , I ,
o 2000 4000 6000 8000 10000

H [A/m ]
Fig. 3.35 B-H curve
A = A (x,y)k (3.237)
In dielectric material potential A fulfils Laplace's equation
V 2A = 0 (3.238)
and in coils satisfies Poisson's equation
V A
2
= -1101 (3.239)
The influence of the beam on the magnetic field distribution may be
modelled by:
• assuming that inside the beam Helmholtz's equation is satisfied:

V 2A = a 2A (3.240)
• or by a relevant boundary condition.
The boundary condition:
~~ = 0, must be satisfied on the surface of the core (11 = 00, a = 0)

(3.241)
and on the plane of symmetry.
Other parts of the boundary, will be considered as:
Applications 139
Fig. 3.36 Magnetic steady state field distribution
(1) ferromagnetic regions placed in an alternating magnetic field, or

(2) thin conducting layers placed on a ferromagnetic wall.
(l) Assume that a real ferromagnetic material is placed in an alternating

magnetic field. A strong skin-effect will occur, and the eddy current will have
only one component
J =Ik (3.242)
The first Maxwell's equation may be written as
curlJ = -«2H F (3.243)
where
«2 = jrolJ,O,
and H F is the magnetic field strength on the surface of and inside the ferro-
magnetic. The continuity condition of vector H at the air-ferromagnetic
boundary has the form
t=10ms
Fig. 3.37 Transient magnetic field distribution
(0 X curllh = -a 2(0 X HFh = -a 2(0 X H) (3.244)

Taking into account that 0 and k are perpendicular we obtain
grad(o· I) = (0 grad) I + 0 X curil =0 (3.245)
After multiplying eqn (3.245) by 0 we obtain
(0 X grad(o . I»r = «ograd)(o x I»r + (0 x (0 x curll»r
= (:n (0 x Ih - a 2(0 x (0 x H)l = 0 (3.246)
Taking into account the strong skin effect

(3.247)
From eqn (3.246) and eqn (3.247) it follows that
(0 x Ih = a(o x (0 x H)h (3.248)
Applications 141
~r=700, 0=7tfS/m
tank
core
shield
beam
~r=700
0=7tfS/ m
~-> ..
~->oo
LV HV 0=0
0=0
eu 4mm
Fig. 3.38 Cross-section of one quarter of a transformer
I
A=Ak divA = 0, H
-
= -curiA
Ilo -
and identity (3.245) we obtain
a(nx (n x H))r =-a (n x (n x curlA))r

ex
= -- {n x «n· V)A)}r
Ilo Ilo
= --
Ilo
"OAJ
a ( (n x k)-
on r
(3.249)
Inside the ferromagnetic material, the vector potential A p satisfies

Helmholtz's equation:
V 2A p = a 2A p (3.250)
and J is calculated from the formula
J = -jc.ooAp (3.251)
Taking into account continuity of A, eqn (3.248) will be transformed into
2
(n x Jh "
= -jc.oo«n x k)Ap)r a «n x k)Ah
= -- "
(3.252)
Il
From eqns (3.249) and (3.252) we obtain
( OA
an
_ PAJ r = 0 (3.253)
where
The above relationship expressed the third boundary condition (Hankel's) and
is very useful for FEM applications.
(2) Thin layer models simplify calculation of complicated electromagnetic

field problems. The structure can be considered as a thin layer if:
• its dimension, in the field penetration direction, is small, compared with

other dimensions,
• a weak skin-effect occurs in the domain.
Theoretically, it is assumed that the domain is an infinitely thin conducting

region with a -+ 00 and it is possible to introduce surface conductivity
as = lim
d-O
do (3.254)
0-->00
which has a limited numerical value.

Let us discuss the model presented in Fig. 3.39.
[2
z
J.1=00
<EE3f-----
0=0
Fig. 3.39 Thin layer model

Applications 143
A thin conducting layer is placed on the ferromagnetic material. The

induced current flows in the z-direction and on the surface f. the following
equation is satisfied
(3.255)
where:
H magnetic field strength on the left side of boundary r I'
HF magnetic field strength on the right side of boundary r h
K surface current induced in thin layer.
Because
(3.256)
then
(0 x H)r\ = (Kh 1
(3.257)
After substitution and transformation we obtain
(0 X Hh1 =-1 (0 x curl A)r, = --1 ((0· grad)Ah = --1 (MAJ

l
-;-k
110 110 110 un rl
(3.258)
Fig. 3.40 (a) FEM mesh and (b) magnetic field distribution in a transformer leakage area
when the beam is modelled by Helmholtz's equation
(a) (b)
Fig. 3.41 (a) FEM mesh and (b) magnetic field distribution in a transformer leakage area
when the beam is modelled by the boundary condition of the third type
The surface current density vector is calculated from the formula

(JS,)f. = (-jroaAh, (3.259)
From eqns (3.257), (3.258) and (3.259) we obtain
(aA
an
_ PIA] = 0
f.
(3.260)
where
PI = jro~oas = jro~oa d
Equation (3.260) expresses the boundary condition of the third type.
We assume that Helmholtz's equation is satisfied in the beam. The problem
formulated above was solved by the 'SONMAP' code. Results of calculations
are presented in Figs 3.40 and 3.41.
4
Reluctance Networks
Janusz Turowski
4.1 A SURVEY OF EXISTING SOLUTIONS AND PROGRAMS

The method of equivalent reluctance networks (RNM) is based on Ohm's
Law for magnetic circuits
VII; = RII;(f); (4.1)
and magnetic Kirchhofrs Laws, for nodes
n
.L; (f); =0 (4.2)
;= 1
and meshes
(4.3)
It is one of the oldest methods of modelling and calculating magnetic circuits

in electrical machines and transformers. As such circuits are becoming increas-
ingly complicated, for example in stepping motors or switched reluctance
motors, the equivalent networks have been developed into extensive multi-
node systems. Moreover, in order to model alternating magnetic fields in the
presence of solid metallic bodies, in particular in iron with its nonlinear
magnetic characteristics and induced eddy-currents, new complex reluctances
have been introduced. The main benefits of using the RNM are evident when
calculating three-dimensional fields of complicated geometries, as significant
savings in computing time and effort may be achieved owing to the efficiency
of the formulation.
The RNM was first used by the author of this Chapter in 1960 [4.22] for
two-dimensional (20) field modelling and calculation of magnetic fluxes in a
three-winding power transformer. In 1969 the author introduced a multi-node
three-dimensional reluctance network model 'RNM-30' (Fig.4.1) for fast
approximate calculations of leakage fields in a three-phase power transformer.
The RNM-30 was presented and discussed at various international trans-
former conferences, including CIGRE'81 USA, 3DMAG'89 Japan, ITMA'91
India, Shenyang'91 China, ISST'93 Poland and TRAFOTECH'94 India, and
continues to provide the basis for many scientific and technical computations
[4.20]; [4.22-4.29].
146 Reluctance Networks
16
~ 0'"" -, /JJ/II .15
14
Fig. 4.1 The basic equivalent reluctance network model RNM-3D for three-dimensional
leakage fields in three-phase power transformers (one quarter of a transformer is shown)
[4.22; 4.26J
The equivalent reluctance network method was developed in parallel by

many other authors. Since 1966 a series of fundamental papers by Davey and
King (e.g. [4.7; 4.15]) and since 1975 works of Carpenter [4.3], Djurovic and
Carpenter [4.9], and Djurovic and Monson [4.10], have contributed signifi-
cantly to the theory of the RNM. Of particular interest is the paper by King
[4.15], who demonstrated an equivalence between the reluctance network
formulation and the finite difference method, and Carpenter [4.4], who
showed similar equivalence with the finite element method. The latter has also
been confirmed computationally using an example transformer by Komtza
et al. [4.16]. In 1979 Kubusch [4.17] effectively applied the equivalent
reluctance model to the computation of leakage fields and stray losses in tank
walls of power transformers. Since 1980 Anuszczyk [4.2] has been using the
two-dimensional version of the method to calculate rotational magnetization
of laminated iron cores of electrical machines, especially induction motors.
Other successful implementations of the method in recent years include the
computation of transient fields, inductances and forces in a linear reversible
electromagnetic pump [4.20; 4.29]. Owing to rapid computation the RNM is
particularly suitable for analysis of transient processes, which often require
a large number of repetitive calculations. The efficiency and good convergence
of the method is demonstrated in references [4.26] and [4.29].
A Survey oj Existing Solutions and Programs 147
The concept of tubes and slices put forward by Hammond and Sykulski
[4.13J, and introduced in Chapter 1, Sections 1.8 to 1.10, is somewhat similar
to the idea of reluctance networks, but takes advantage of the inherent duality
of electromagnetic fields. At present the RNM uses one solution, although
provision of dual bounds for this method could be one of the possible future
developments of both techniques.
A series of papers published by the author of this Chapter and his
collaborators during the past decade [4.20; 4.25-4.29J have concentrated on
developing a convenient engineering computational tool for field modelling
and optimization of leakage zones of large power transformers. For such a
tool to be acceptable for regular design usage, a number of demanding
requirements must be met.
• The program must be easy to use and perform satisfactorily even on low-
cost personal computers.
• Computation has to be very fast and interactive, even in three-
dimensional fields.
• It must provide useful design parameters.
• Flexibility must be maintained with minimum cost in time and effort.
• The program must allow the incorporation of complicated, three-
dimensional, three-phase geometries, magnetic non-linearities, heating
effects, deviation of material parameters, and other important practical
considerations.
Although, in general, the application of electromagnetic software for design

purposes is rapidly increasing (see the discussion in Chapter 7 of this book),
for some more complicated devices, such as power transformers, unreasonably
high computing times and memory requirements make it too expensive and
impractical for standard packages to be used as every-day design tools. Many
design and maintenance engineers would argue that the main objective is not
to calculate the field with the highest possible accuracy but to quickly, and
at an early design stage, localize and eliminate excessive stray losses, crushing
forces or overheating hazards. Clearly, compromises have to be made in terms
of accuracy or even adequacy of the model. Approximations, simplifications
and even rough estimates may have to be introduced. The expertise of the
designer becomes invaluable, and experimental verification of the model
becomes part of the procedure.
The RNM-3D package offers such a simple, user-friendly, inexpensive
engineering tool. The designer is provided with an answer within 15-30 s for
each computational variant, which is a clear advantage in the design environ-
ment. A dedicated pre- and post-processor is built in for a selected class of
three-phase transformers, and a general purpose solver is available (see
Fig. 4.12) for solution of multi-node electric networks. In order to model
other types or classes of electrical devices, say actuators or rotating machines,
the pre- and post-processor has to be modified in accordance with particular
requirements. A library of typical geometries, problems or devices could thus
be created. Examples of such successful applications include calculation of

three-dimensional fields, forces and losses in end-windings of large turbo-
generators, by Davey and King [4.7]. Two-dimensional transient fields have
been simulated amongst others by Oberretl [4.18] and Demenko [4.8] using
equivalent RC networks with mutual capacitances.
4.2 RELUCTANCE NETWORKS AND FINITE DIFFERENCES

The method of equivalent reluctance networks has been extensivly reported,
starting with the first paper by King [4.15], in which a theory of the RNM
is derived from the well-known method of finite differences (see Chapter 2
for a full discussion of this technique). For example, Poisson's equations for
the magnetic vector potential A in two-dimensions, with the components of
current density J=J.. ,Jx=Jy=O, and thus A=A.. =A, Ax=Ay=O, at
constant permeability 11,
a2A a2A
ax 2 + ay 2 = -IlJ (4.4)
can be transformed into a difference equation by expanding the function

A (x, y) into a Taylor's series along both axes x and y (Fig. 4.2), where the grid
with a mesh dimensions a x h is rectangular and can be adjusted to fit a
particular shape.
Assuming the coordinates of nodes 1 and 3 to be XI = X + hand X3 = x - h
respectively, we find
h h2 hn
AI =A(x+ h) =A(x) +-A'(x) +-A"(x) + ... +-A(n)(x)
I! 2! n! (4.5)
h h 2 hn
A 3 =A(x-h) =A(x) --A'(x) +-A"(x) - ... +-A(n)(x)
l! 2! n!
Adding both sides of eqn (4.5) and assuming A(x) = A o gives
a2A(x,y) = A "()
x = h2
--a":"x~2":"":"
1 [
AI + A 3 - 2Ao - E
(
h
4)] (4.6)
Fig. 4.2 A rectangular element of the two-dimensional reluctance grid

Reluctance Networks and Finite Differences 149
By analogy, for nodes 2 and 4 with coordinates Y2 =Y +a and Y4 =Y - a,

and taking A(y) =A o, yields
2
a A(x,y)
ay 2 = A" ()
Y =2
1 [
A2 + A4 - ( 4)]
2A o - E a (4.7)
a
The last terms on the right-hand side of eqns (4.6) and (4.7) indicate the order
of the error of approximation. Neglecting these errors leads to the following
relationships
a2A AI A3 2A o a2A A2 A4 2A o
ax 2 == h2 + h2 - -,;r- and ay2 == a2 + a2 - (;'2" (4.8)
Hence eqn (4.4) takes the finite difference form
Multiplying both sides of eqn (4.9) by ah/IJ. and introducing values of

elementary reluctances per 1 metre of length along the z axis (see Fig. 4.2) in
the following form
Rx=_h- and R =_a_ (4.10)

lJ.a·1 y IJ.h· 1
we obtain
(AI - Ao)Ry + (A 3 - Ao)Ry + (A 2 - Ao)Rx + (A 4 - Ao)Rx = -10
(4.11)
where 10 = Jah is the total current in a single mesh.

Considering now the relationships
<1»; = IIB; . dS; = IIcurl A; . dS; = ; A;' dl;

Sj Sj
(4.12)
we can write for a single mesh (see Fig. 4.3)

<I»;=Ao-A i , where i=I,2,3,4 (4.13)
Fig. 4.3 Determination of mesh fluxes +; = EA;I; with the help of the magnetic vector
potential A; using eqn (4.12)
Fig. 4.4 Equivalent reluctance circuit for a grid of Fig. 4.2
and then, after substituting eqn (4.13) into eqn (4.11), we obtain Kirchhoffs
and Ohm's equations for magnetic circuits
4
<l>tRy + <l>3Ry + <l>2Rx + <l>4Rx = 10 or ~ ;R; = 10 (4.14)

;= t
for a corresponding equivalent magnetic circuit (Fig. 4.4).

The problem has now been reduced to setting up an appropriate grid model
corresponding to the field in a specified device, and then solving it using any
available program for multi-branched electrical networks, such as NAP-2,
MIMIC, SPICE-2, or NODAL. There are two particular requirements which
must be carefully investigated before the solution is attempted. First, the
program must be able to handle a larger number of nodes, typically 500 to
1000 or more, and secondly, the solution times offered by the software must
be extremely short, of the order of seconds, in order to make the process a
fully interactive session, even on a personal computer.
4.3 LEAKAGE FIELDS IN TRANSFORMERS

In order to illustrate the method, and demonstrate its effectiveness for a
typical important design problem, we shall look now at a model of a leakage
field in a power transformer (see Fig. 4.5).
Depending on the accuracy required, various grid densities may be used. In
the following examples most simulations have been performed on a grid with
about 90 nodes (see Fig.4.5b), but some of the preliminary calculations
have also been made on a grid with about 50 nodes (see Fig. 4.5a). The
argument here is, as demonstrated in Fig. 4.6a, that the specifics of the trans-
former structure make the convergence of the solution, with regard to the
number of nodes, relatively high, and a coarse grid is often sufficient for
preliminary assessment of results. For instance, the calculated distributions of
the radial component of flux density in the Low Voltage (BLV ) and High
Voltage (BHV ) windings, as well as the tangential component of field intensity
(HS,T) on the internal surface of the tank wall (Fig.4.6a), with 46 nodes
Leakage Fields in Transformers 151
(a) t=:!!::=iJiI;;.;;-::-"C J- ;t,-C:' J-.';-c:':l-7- 24

II f I ,'23
.......-d!lllll........:c:J-~-e:J--X
I f-C:::r-4-c=-'''''.... :>-1,., 20
"y
I
I 163
w
I:t:
2I+~
z
I o
0
141
14 ~ U 3J~
W
0::: Z
15
I o 10
6
~
12
I 4
8
I
@
~"_.
2
• l! "
"
10 8 6 5 4 "3 0 14 12 tl8 6 2 0
Fig. 4.5 A reluctance network model of leakage field in a power transformer (a quarter of
the cross-section is shown) f4.25}; (a) with 46 or 51 nodes, (b) with 89 or 91 nodes
(a) IX
lX
B~V
(b)
so
1~ 3
200 .?~ c('(Hol
fe) t perm t I2ato'
Fig. 4.6 The RNM-2D analysis of leakage fields f4.25; 4.26]: (a) effects of number of
nodes, (b) hot-spot localization at permitted temperature tperm and corresponding permitted
field Hperm = 40A cm- I
(dashed lines) and with 89 nodes (continuous lines), are practically the
same.
A good accuracy of the network model of the leakage field of transformers,
even for a small (50-100) number of nodes (Fig. 4.5), is not surprising if it
is remembered that the successful classical method of leakage field calculations
is based on a single reluctance of the inter-winding gap, corrected only by the
Rogowski's coefficient [4.22; 4.26].
Currently programs with a much higher number of nodes, up to 1000 and
more, are available. However, increasing the number of nodes beyond what
is reasonably required is not recommended, as a slight improvement in
accuracy may not be a sufficient justification for increased computing times,
and thus the method's main strength, its fast response time, could be impaired.
Therefore, since speed is of the essence, to make the computation faster,
whenever possible analytical formulae should be used for computing
particular reluctances, as well as for recalculating parameters at the post-
processing stage.
In order to illustrate the computation of grid parameters (reluctances) and
to facilitate understanding of this process, the example from Fig. 4.5a will be
used, with a relatively small number of nodes. The discussion below effectively
defines the way in which an analytical pre- and post-processing is created for
the interactive program RNM-3D (Fig. 4.20). The example of Figs 4.5 and 4.6
is taken for a 2D plane at y = O.
It follows from Fig. 4.2 that the elementary reluctances (in 1/H) given by
[.
Rlii=----!....-S' i = x,Z (4.15)
J.l.o i
are calculated from geometric dimensions of meshes (Fig. 4.7), into which the
investigated area is subdivided. The number of branches of the network is
determined exactly by its discrete coordinates i =xyz of the cartesian
coordinate system (Figs 4.1 and 4.5a). For every reluctance in Fig. 4.5 there
are corresponding magnetic flux components ~x and ~z passing through this
element. All quantities Vxz = Rxz~xz, Rxz and ~xz are uniquely localized in the
network with the aid of indices xz corresponding to discrete coordinates of
the system. Complex reluctances Rxo of the solid iron tank walls and cover
contain additional nodes, brought about by the subdivision into real and
imaginary components owing to the eddy current reaction.
Determination of magnetomotive forces of sources

Owing to the slenderness of typical windings of large power transformers, in
many cases, especially when the field outside the windings is investigated, the
magnetomotive force (mmf) of the windings can be evaluated by assuming
that entire current flow of the High Voltage (HV) and Low Voltage (LV)
windings
(4.16)
Fig. 4.7 The principle of calculation of elementary reluctances of a network: R x is in the

x-direction and R z in the z-direction
lIy i
-----------------------~
,
r- --- -- _oJ
pr, ,
I
I t-e::rt-c:J~
I •
I
plj
L_ H=>+<~
__________________ J
(a) (b)
Fig. 4.8 Modelling of the magnetic circuit of a reactor using the RNM ([4.22J, p. 268):
(a) the cross-section of a wound central leg, (b) the equivalent network.
R L reluctance of the central magnetic leg
Ryj reluctance of the ith section of the yoke
o i&I actual currents
ElI8J fictitious currents
11 mmf compensating the fictitious currents
is concentrated in a filament layer (current sheet) on the axes of the HV or

LV windings.
If we wish to account for finite dimensions of winding cross-sections, we
can use a well known principle by which the field of a current flow, subdivided
into elements l1ij N j (see Fig. 4.8), is replaced by a network of circuits. This
procedure consists of the shifting of the elementary current l1ij N j from one
mesh, e.g. (x = 3, Z = 1), across the border between the nodes 22-42, to
another mesh, say (x = 3, Z = 3). This produces in this border a new fictitious
mmf source with the value of shifted flow F = l1i(N(. At the same time the
flow in the mesh (x = 3, Z = 3) is doubled. As the result the first mesh
(x = 3, Z = 1) becomes empty. This can be explained by postulating that in the

axis of the border between adjacent meshes we can insert a fictitious coil
(rectangles in Fig. 4.8), with identical but opposite flow (-M 1N() from the
first mesh, and simultaneously compensate the action of this coil with the help
of a fictitious mmf of the same but oppositely directed flow. This implies an
additional rule - that every shifting of ampere-turns !1i;N; across any mesh
border creates an mmf (flow) F = M;N; in the axis of this border and doubles
the flow in the adjacent mesh. The next shifting of ampere-turns follows
the same rule, but this time with the entire enlarged mesh current flow - Le.
its own and added.
By repeating this operation we can replace a real coil with a distributed flow
(Fig. 4.8a) with a circuit network of reluctances and lumped magnetomotive
forces (Fig. 4.8b). In an analogous way one can shift the current flow to a yoke
axis.
In the case of height asymmetry of HV and LV windings (Fig. 4.9), an
additional transverse flow Fz (Fig. 4.9) should be introduced. This radial flow
can be found using a well-known rule of ampere-turn decomposition into axial
and radial components ([4.22], page 347). In this way any asymmetry can be
considered, with the accuracy depending on the number of nodes introduced.
Because of the slenderness of the windings of large transformers and,
usually, the asymmetry of HV and LV windings, we shall limit our attention
to the flow F calculated using eqn (4.16). Calculations for higher numbers of
nodes involve evaluation of submultiples of the given values [4.25].
Evaluation of network elements

The elementary mmf Up of an inter-winding gap for a bigger mesh (Fig. 4.5a)
can be expressed using the formula
=--- .fiIHNH = 1 per umt

n1 =
Up F . va Iue (b' .)
aSlC umt (4.17)
n1
n+- n+-
2 2
INM-F
h z F = V2'1 N!!.
z h
:(
I:
I IN IN
~
2; 'g:j
""1 0 e 0 <:9 + <G o
8 o
(a) (b)
Fig. 4.9 The influence of winding asymmetry on the RNM model [4.27/: (a) decomposition
of ampere-turns into a symmetrical component F = J2/N and a radial component with
maximum flow F = J2/N M h -I, (b) equivalent reluctance network with addition source Fz in
the radial branch, where Fx is an element of the main flow
where n =6 is the number of bigger meshes along the entire winding height,
nl = 4 is the number of smaller meshes on both edges of the winding, and
p = 14, 34, 54 denotes coordinates (numbers) of branches referred to by
eqn(4.17).
The elementary mmf of smaller meshes at the edges of windings (Fig. 4.5a)
are given by
Up
Up1=T (pi = 74, 94) (4.18)
The reluctances of a leakage zone of a transformer are determined from its

geometric dimensions (Fig.4.10). In order to facilitate the calculations, the
reluctances have been expressed in relative units with reference to the
reluctance of the equivalent inter-winding gap Of, for bigger meshes.
As an example, the reluctances of the inter-winding gap and above it
(Fig.4.5a, 4.10, and 4.11a) are:
(a) for bigger meshes

1 hR
R =- - -----"------ = 1 per unit value (basic unit) (4.19)
p nl s: , Imean
n + 2" lJ.ou -s-
Fig. 4.10 Geometrical data necessary for calculation of reluctances in the RNM-3D model
[4.26J
(f)
(d)lll~ (e) 1~
OIL
niriTt z~
<
f-<
LLl·····_·· ~
jill kcl ~ l) CD
a::
8 - - - - - - --
11
IIII U IPf4'/1 R.r iJ:!Jfi ll ·<r=O

o
Fig. 4.11 Computation of reluctances: (a) axial along the gap axis, (b) radial towards the tank, (c) radial towards the core, (d) axial between the
tank and the winding, (e) complex, of the tank wall and cover, (f) electromagnetic copper screen -
where p = 14, 34, 54 denotes coordinates (numbers) of reluctances, and s = 16

is the assumed number of segments on the leg circumference (Fig.4.11a):
0'
R p2 =Rp - - - - (p2 = 154, 174, 194, 214) (4.20)
0' + 02 + e'
2
e =! (e + e lV )
H lII
hmean = ! (h uL + huH)
0' = 0 +!(al + 02)
1
hR = "K hu
K =1_ a, + 02 + 0
7th
Imean = ! (11m + 12m )
(b) for meshes half as small the reluctances are half as small
RpI = ! Rp = 0.5 per unit value (pI = 74, 94) (4.20a)
1 0'
R p3 = 2. R p (p2 = 114, 134) (4.20b)
0' + 02 + e'
2
The radial reluctances towards the tank wall (Figs 4.5a and 4.11 b) in 1/Hare
(a) for bigger meshes
(4.21)
(b) for smaller meshes these reluctances are correspondingly increased

R kl = 1R k and R k2 = 2Rk (4.21a)
Other reluctances are (Fig. 4.11):
• radial to the core R, (Fig. 4.11c), with coordinates r = 25, 45, 165, 185,
205, 225, ,1 = 65, 145 and '2 = 85, 105, 125;
• axial along the oil gap between the tank and winding R x (Fig. 4. lid),
with coordinates x = 12, 32, 52, 152, 172, 192, 212, xl =72, 92, 112,
132, 238, 231;
• complex, axial reluctances (or rather magnetic impedances) for conduc-
tive areas Rs = R sR + jRsL of the tank wall and cover, made of solid
steel (Fig. 4.11e), with coordinates s = 10, 30, 50, 150, 170, sl = 70,90,
110,130 and s2 = 191,213,235,239. These reluctances take into account
analytically the skin effect, eddy current reactions with phase shift,
nonlinear permeability inside solid metals, equivalent depth of an
alternating field penetration into solid metal and the screening effect
of a double-layer of the screen and solid metal (Fig. 4.11 f).
The unitary reluctances of the solid steel wall have been calculated as
complex values with the help of the formulae ([4.22], p. 139)
(4.22)
where
R~I = al~(:~J and R~lr = a2~(:~J (4.23)
and al'" 0.37 and a2'"

0.61 are linearization coefficients for solid steel
(al a2
= = 1 for non-magnetic metal).
The linearized formulae (4.23) are valid for solid metal walls of thickness
d larger than the half wave length A. in the metal, Le.
d>~, 1..=21t~(ro~~J (4.24)

In the case of solid steel the threshold value of 1../2 is about 4-5 mm, whereas
a typical tank wall and cover of a large transformers have a thickness of
10-15 mm or more. Also, ~s = ~O~rs> where ~rs is taken from the magnetiza-
tion characteristics ~ = ~(H) ([4.22], pp. 68, 71) for the tangential H ms on the
internal surface of the steel wall.
Electromagnetic screens of the tank wall

If the tank wall is covered with an electromagnetic screen (e.g. made of copper
or aluminium), the complex reluctance of such a conducting wall in per unit
values becomes ([4.27], p. 170)
R~ _ a,a,siaha,+ + a,d~~J]
-el ~oa3[1 + a2d~(:::J] (coshazd-l) + ~wazsinhazd
a2 sinh azd 2
(4.25)
,.. ~o(cosha2d - 1) ,.. ~od
for ladl z <c 1, where dis the screen thickness, aj= (l + j)kj, kj = (1tf~jaj)-II2,
i = 2 for the screen and i = 3 for solid steel.
For example, for a screen with d = 4 mm and f = 50 Hz, kzd = 0.428 and
Rei'" 0.4 X 109 H- 1• This means that with an accuracy sufficient for engi-
neering practice, one can assume, as a first approximation, that in the parts
of the tank covered by an electromagnetic screen made of copper or
aluminium, the tank reluctance is infinitely large, i.e. R s = 00.
Magnetic screens (shunts) on the tank wall

If the tank wall is covered with a magnetic screen (shunt), made of laminated
iron, the resultant reluctance Bs (in H -I) of the solid steel and the screen can
be calculated (see reference [4.22], p. 172) as a parallel connection of the
reluctance Bs of the screen and Bw of the solid steel wall:
B s =- - - = - - - - - 7 71" " - - - : : -
_1 + _1 ~ d+ 1 - j f[2~wJ + n, ~J 1t(D + 2ac + 2aT)
Be Bw e 2 'V roo
(n
(4.26)
where ~e and d are the magnetic permeability and thickness of the screen,
and 0 and ~w are the conductivity and magnetic permeability of the tank wall.
The minimum thickness, with regard to sheets saturation, of the magnetic
screen can be evaluated (see reference [4.22], p. 197) from the formula
0c .fiIN~oo'
d min ~ (4.27)
ac + aT B'f.XX)h
where B,f.XX) is the maximum flux density corresponding to the screen
permeability ~er = 1000 (i.e. about 1.7 T for anisotropic and 1.4 T for
isotropic transformer steel). Neglecting the reluctance of the solid steel wall
provides an additional safety margin for the evaluation of dmin •
The reluctance. of the steel wall covered with an unsaturated magnetic screen
is so small compared with the reluctances of nonmagnetic regions, that one
can safely assume, as a first approximation, that this reluctance is zero
(4.28)
whereas for a non-screened steel wall and cover, the value of R s1 has to be
calculated using eqn (4.22).
All the other reluctances of the three-dimensional model (Fig. 4.1),
including those in the circumferential direction (y axis), have been calculated
as explained above. The detailed formulae are given in reference [4.27],
Chapter 3. The appropriate values of these elements, corresponding to the
declared transformer parameters (Fig. 4.10) and the frequency of the supply
voltage, are placed in the network scheme, along with the voltage sources
which model the elementary magnetomotive force in a winding for each phase.
All network elements are calculated automatically in a preprocessor on the
basis of the analytical formulae. They are usually expressed as per-unit
(relative) values referred to the data of the inter-winding gap o. The super-
script 'r' means a relative value referred to the gap value.
The nonlinearity of the permeability of steel with depth has been considered
with the help of linearization coefficients ap '" 1.4 for resistance and active
power, and op'" 0.85 for reactance and reactive power ([4.22], p.322). The
permeability along the x axis has been taken as constant in the first instant
but it is possible to take into account the changes of ~ using an iterative
method or an analytical approximation.
We shall look now at the post-processor of the program RNM-3D. As a
result of the computation, two quantities for each reluctance are of particular
interest, the flux density B mi and the magnetic field strength H mi • Hence
Cl>mi Vmi
B mi =T I
and H mi =T
I
(4.29)
where Cl>mi is the maximum value of the flux calculated for the ith reluctance,
in Wb, Vmi is the maximum value of the magnetic voltage, in A, and Ai and
Ii are the cross-section and length of the ith reluctance, in m2 and m
respectively.
The flux density B mi gives rise to forces and additional losses in the
windings, whereas the tangential value of H msi on the steel surface can be
related to induced eddy-current losses in the tank wall and cover, yoke beams,
clamping plates etc., which are often calculated using the following formula
([4.22], pp. 193, 330).
PI = k(d)! ~[~:oJ H~(~'S) IH ms l

2
dS (4.30)
s
where k(d) depends on the wall thickness, k(d) < I at d < 1../2 and k(d) = I
for d ~ 1..12; ap '" 1.4 is the linearization coefficient ([4.22] pp. 322-323) for
solid steel, ap '" I is a linearization coefficient for a non-magnetic material;
~rs = ~rs(Hms) is the surface relative permeability, ~s = ~O~rs; (J is the
conductivity and A = 21t(ro~(JI2)-1/2.
When the calculated quantities are expressed in the per-unit notation as
CI>{"i' B{"i and H{"i' the corresponding reference values for the inter-winding
gap 0 I, for a sector of width 21t1s (see Fig. 4.11 b), are given by
A..
"VmS -
_ J2/hNH N H ~ou~ -s-
Imean --
I I . val ue (Wb)
per-umt (4.31)
R
Bms = J2/hNHN H ~o = 1 •
per-umt value (T) (4.32)
R
H = J2/NHN H = =I per-unit value (Am-I) (4.33)

[nlJ
S VmS
hR
m
hR n +"2
where /NH is the rated current in the HV winding, N H is the number of turns
of the HV winding, and both are phase values. For other symbols refer to
Fig. 4.10.
DATA . . . - - - - - - , SOLUTION
PREPROCESOR POSTPROCESOR
INTERACTION
Fig. 4.12 Interactive design using the RNM-3D package [4.26J

Computational Aspects 161
4.4 COMPUTATIONAL ASPECTS

Having formulated the reluctance model for a given device, or its part, see
Fig. 4.1 or 4.5, and calculated all the network parameters using the formulae
given in Section 4.3, or in reference [4.27], p. 151, the next step is the solution
of the multi-mesh network. This can be done using one of many popular
programs for network analysis, such as NAP-2, SPICE-2, MIMIC, NODAL
etc. It has been found, however, that it may be advantageous to create a
dedicated solving routine, in order to benefit fully from the particular features
of the method. Thus a fully integrated system has been developed for fast
interactive analysis of the leakage zone of power transformers. The package
is called RNM-3D and incorporates the pre- and post-processing as described
in Section 4.3 (Fig. 4.12).
The flowchart of the RNM-3D is shown in Fig. 4.13. The basic reluctance
model was shown previously as Fig. 4.1. Different variants of the model are
stored on a disk. Every variant provides a skeleton framework for calculations
and requires actual reluctances and some other data to be specified for each
particular case. A node admittance matrix is then created for the network
solution, using the modified nodal approach [4.14], and a set of linear
equations is solved using a direct sparse matrix method [4.11]. Finally, post-
processing is applied and distributions of the magnetic field along selected
paths are displayed or printed, as shown in Figs 4.6, 4.14, 4.15, 4.16 etc. For
running the simulation and analysis of a three-phase transformer, the user may
employ the existing network model (Fig. 4.1), stored on a disk, or define his
own model of the same topology, but with a different number of loops and
branches. The user's model may also be saved on disk for later use.
The main program (Fig. 4.13) must be provided with the following input
data:
(1) a 'PARAM' statement, followed by some constructional parameters of

a transformer:
• the maximum current I Hv (IHV) in amperes,
• the number of turns ZHV (ZHV),
• the following dimensions in millimetres (see Fig. 4.10): 01 (AI), 02
(A2), 0 (DEL), hLV (HLV), hHV (HHV), D (D), c! (Cl), CIl,JlI,lV,V
(Cl, C3, C4, C5), M (M), dAr (DAL), d eu (DCU), h k (HK), hp
(HP), hd (HD), b l (Bl), b l ' (BIPR), b2 (B2),
• the relative magnetic permeability of steel J.1rs (MRS) - if neces-
sary obtained through iteration from the magnetization curve,
Fig. 4.17b, or from the analytical approximation, eqn (4.35);
(2) an 'F' statement, followed by the value of the supply frequency or the
range of frequencies and the increment;
(3) a 'MODEL' statement, followed by the model number which is to be
retrieved from the disk for analysis:
An example set of data may be found in paper [4.26]. Following the

instruction 'RUN' , the calculation of parameters is executed and the results
( START )
T
/ Read data: START. PARM-l. F. MODEL.Rememberc.om~ /
r •
Calculate the values of network relucloncl?s I
/
•
Read: MODEL. CHANGE. SCREEN. BEAM /
~
/ Read -RUN /-
.~
1 2
l-
~
?
not changed
~ t
r iJke the netwo.1< !opolo9icol data (names 01 elements.~ runbersl from the ex!l'm:llmemary I
1
r Put the celculated reluctance values to the propt'r elements of the network I
NO~
Yes
r Exchonqe the values of indicated network elements . I
~
No
Yes
I Cclculate new 'oClues d Ri • Z{ at the pbce of scrfffl and put it in\::l the Ol'twork I
~
No
? Yes
r Calculate new values of R, at the place of. beam and put it into the netwo.1nts in the specified brooches d the ~rk I
T
r Calculate Ihe maqnetic values sf and Hi in the specified bronches I
t
r Display (print) the rel;ults.tables and diaqrams I
~
No
Yes
( STOP ')
Fig. 4.13 Flowchart of the RNM-3D algorithm [4.26]
Computational Aspects 163
lOCAl. tJY[IIlAIING
/lA/AID OftQRE
Fig. 4.14 The RNM-2D analysis of the influence of a laminated magnetic shunt collector
on the magnetic field as in Fig. 4.6a [4.25}
<r
w H----,---.
~
u
d
c
r
H
s ,.T
O'1)'~&7egIC
-2
xl0 p. u. value
Fig. 4.15 The RNM-2D analysis of excessive heating hazard of flanged bolt joints with
regard to their position on the tank wall ([4.27}, pp. 188-194)
are displayed. The reluctance network thus created can be modified using the
following special commands:
• 'CHANGE' - for changing the values of specific elements of the

network. The elements can be identified by their position in the xyz
coordinates of the network (Fig. 4.1),
• 'BEAM' - for the simulation of a horizontal solid-steel yoke beam or
a laminated magnetic shunt collector (Fig. 4.14) above the transformer
windings,
• 'SCREEN' - for the simulation of copper or aluminium screening plates,
of specified thickness, on the tank wall.
Other modifications can also be made. For example, the designer can
specify coordinates of the reluctances which are to model the presence of a
copper screen or a magnetic shunt (see Fig. 4.16). The five-limb transformer
can be represented too, by specifying the reluctances with coordinates (x = 1,
3, 5, .. 0' 21; y = 0, z = 2) and (x = 20, 22; y = 0; z = 1, 3), as well as those
on the opposite side of the core, as equal to zero.
3
2
'--=--4-:'-~S- -2 - HT
xlO p.u.
(a)
(b)
Fig. 4.16 The RNM-2D analysis of the influence of screens of the tank wall on the field
H msT on the tank surface: (a) electromagnetic (copper) screens, (b) magnetic (shunt) screens
made of laminated iron
4.5 INTERACTIVE DESIGN AND DETECTION OF

'HOT SPOTS'
Most of the electromagnetic software used currently to aid the design of
electromechanical devices utilizes field analysis, with prescribed boundary
conditions and excitation, as the main computational tool. Field synthesis,
although fast developing (see Chapter 5) and already capable of dealing with
simple geometries and problems, has not yet reached the stage of a fully
matured design tool, and very few applications to real engineering problems
exist. The most promising, for some time to come, is the idea of an interactive
design, where the designer is in constant 'conversation' with the computer.
Although some elements of the process may be automated, and proper
optimization techniques are under development (see Chapter 7), it remains
essentially a trial-and-error approach, with the designer making all the
important decisions and the computer, with its suitable software, facilitating
Interactive Design and Detection of 'Hot Spots' 165
the computation and providing useful post-viewing and post-processing. Fast

computation becomes essential and many techniques, including the most
popular finite-element method, may, especially for the more complicated
three-dimensional problems, require far too long computing times and thus
become impractical for the interactive process. The dedicated approximate
computation, with its particularly short solution times as offered by the
reluctance network method, becomes an interesting alternative, especially
for simplified engineering analysis and design. For example the RNM-3D
program provides an approximate answer typically within 15-30 s for one
computational variant of a transformer leakage field. This should be com-
pared with possibly even hours of computation using finite elements at this
level of complication.
Before we proceed with further discussion of the RNM-3D, it is worthwhile
to mention an interesting hybrid approach, proposed by the author (Fig. 4.17),
which mixes finite elements and reluctance networks. Thus solid steel elements
(tank and cover) are represented in the finite element model by equivalent
reluctances of laminated (non-conducting) steel of thickness equal to the skin
depth of the element.
Computation of losses in solid metallic elements

Equation (4.30) implies that it is the surface value of the magnetic field
strength H ms which governs the value and distribution of local loss density,
and thus local heating and 'hot spots', in the tank wall and other solid steel
elements of a transformer. Hence the problem of loss computation consists
mainly of evaluation of the tangential component of the magnetic field
strength on a metal surface.
Although the experimental verification conducted and reported in reference
[4.26] confirms a satisfactory accuracy of field computation with the
RNM-3D, loss estimation creates new, more difficult problems, and a higher
level of non-linearity, which may have to be studied separately.
Dependence of stray losses on rated power and load of a transformer

The active power loss in Wm- 2 generated in a massive steel 'half-space',
per unit area on its surface, may be expressed ([4.22], p. 152, 328) using the
formula
p
1
=
a
p
frO)~sllHmsl2 =
\, l20j 2
0) f[O)o]
ap \ ' 2~s
IcI>ml1
2
2
= ~ IH
2~2 ms
IIc1> I (4.34)
ml
where
is the field strength on the surface of iron, in A m -I, op = 1.3 to 1.5 (typically
1.4) is a semi-empirical linearization coefficient for solid steel ([4.22], p. 322),
and ~s is the magnetic permeability of the iron surface (Fig. 4.17b).
AIR
GAP
LAMINATED
CORE
I Ax-
•
P
~,,,
CEOH!:TRy or TIW'STtll\I'Dl
• 1.4.; • • 0.85
q 6
: 4oI"'2d"; .-7 .. 10 51.
7
Id"."l_~
a· •
r-
I
I
F ["
X-<tH'Ol/D(T.
IKlIILtllEAR
"axe hcl
~
at z-o
I
I -~ ~-- ~:H
L-
~ -f
ms
H
H (xl"
..
f2 H lIltS
(x)
or
-JO.67) ff
R "Ax. ylO. 37-
- Pt"' r,- - 2 -
2
; ; ; ' I Has I 11Haslx11>"..
PO'"" A
I·~fll"ap w;;:
". J'o1/P\dxdy
·IH.. lxll
1 1
PIll KT 7 AIIlL PIli KT TAIIlL'
T
PIlI KT TAIIlL,
rRIKT 7Am.[
R -Rlx.y)
x.y Pt"Ptlx. y ) ".-It(xo · yo ) f.I"C(xyzl
2
t/Hol \11. A/a lIb/.
Tota. P. W
Fig. 4.17 The hybrid finite-element and reluctance-network approach for modelling
magnetic fields in transformers or other devices with solid steel elements: (a) the finite
element model, (b) computational flowchart [4.27}
Interactive Design and Detection oj 'Hot Spots' 167
Although eqn (4.34) is simplified, using a more accurate mathematical

description may not necessarily lead to better accuracy of results, as the
material data required for computation, in particular conductivity and per-
meability (or B/H curve), often show significant deviation owing to the
varying chemical composition, mechanical and thermal treatment etc., and
thus the computational effort may be hindered by the lack of reliable infor-
mation on material properties. In Fig. 4.18 magnetization curves of various
samples of a common steel St3s at 20°C are shown.
In Fig. 4.18 the approximating curve (a) H= LOS· 10- 5 (H--./IJ.,)I.77 corres-
ponds to (b) HZ.J;, = 6S0H1.565. Introducing into eqn (4.34) the analytical
approximation (Fig. 4.18) of the magnetization curve of steel
Hms = C(.J;,IHmsl)n = c [~(~~:j lmllr (4.3S)
where C = LOS X 10- 5 and n = 1.77, we obtain ([4.22J p.329]) the relation
between power loss in a massive steel element and the field H ms on its surface
(in Wm- z)
p =
I
--.!!L {COIJ.°I
2Clln \J. 20 H ms
11+lln =:: k
I
X 11.6 xjO.5 X 0- 0.5 (4.36)
or
C co I + n/2 on/2
p =-
I 2.J2 a;IJ.~/2
 I II + n
ml
=:: k X
Z .
I Z.8 X fl.9 X 0°·9 (4.37)
The last formulae demonstrate the dependence of losses on physical

parameters and current. The crucial factor here is the way in which the field
is excited. We can distinguish two typical cases:
(I) when H ms is proportional to I (e.g. the field due to the current in the
bushings on the cover plate surface), we should use eqn (4.36), and
(2) when <l>ml is proportional to I (e.g. leakage flux in the inter-winding
gap), then eqn (4.37) holds.
Since in the RNM-3D model (Fig. 4.1) the magnetic field H ms is calculated
from the leakage flux <l>ml at constant iron reluctance, it follows that value
of power in eqn (4.30) should be somewhat higher than 2, say (Hmsix, where
the semi-empirical correction coefficient x may be selected analytically or
experimentally.
The total loss (in W) on the internal surface of the tank wall can be therefore
calculated from the following general expression
p=~~[~IJ.0J [Pe II~IHmsIZdAe+Pm I I~IHmsIZdAm+

°st A e Am
Pst I Iv'ii:IHmsl2xdAst]
A"
(4.38)
f
2 A .r-
xl0 VIJ- H
m r
foo
H
20 JO +0 so to 70 2 A
xl0
m
(a)
A 2
6 A
2 Vi1r H
xl0
2
m
&
"00
H
o 10 1'0 30 s'o 70 2 A
xl0
m
(b)
Fig. 4.18 Analytical approximations for magnetization curves of steel [4.24]:

1 - (00) 0.2C, 0.15Si, 0.03P, 0.5Mn, 0.375, 0.05Cr, 0.03Mo, 0.7W; (] = 7.2' ul S m- I
2 - (x x) 0.16C, 0.25Si, 0.015P, O.013S, 0.59Mn; (] =5.69' J(f S m- I
3 - (M) max 0.22C, (0.1-0.35)Si, max 0.05P, max 0.05S; (]=6.84·J(f Sm- I (Polish
Standard)
4 - (00) 0.3C, hot-rolled
Interactive Design and Detection of 'Hot Spots' 169
24.0KVA Trans{onDcr 127KY... Transformer 31SkVA Trans(orlDcr
Fig. 4.19 The 3D loss density distribution in a quarter of a non-screened tank surface
(Fig. 4.1) calculated with the RNM-3D package [4.24}
where Pe <C 1 and Pm <C 1 are the screening coefficients ([4.22], p. 198 and
2(0) of electromagnetic or laminated magnetic screens, respectively, A e, Am
s,
and A are areas covered by a corresponding screen (e or m) or area which
is not screened (st). The variation of power for f and (} is less important at
typical frequencies and temperatures [4.24]. The value of x varies for different
transformers but is typically between 1.1 and 1.14. This has led to the
following calculated total losses in the tank wall without a screen: 288 kW for
a 240 MY A transformer, 78 kW for a 127 MY A transformer, and 88 W for
a 315 kY A unit (Fig. 4.19).
Bearing in mind the inevitable deviation and non-linearity of material
parameters, the fact that x usually has to be estimated experimentally should
not be considered an unreasonable restriction. The overall accuracy of
calculation in any case cannot be better than the natural spread of electric and
magnetic properties of materials.
Figure 4.20 shows the results of the interactive design of combined copper
(Cu) and magnetic (Fe) screens, including calculation of loss density
distribution and optimization of the design with the view to reducing stray
losses and local heating.
Detection of 'hot-spots' in the leakage region of a transformer

Peak values of field and loss density distributions in Figs 4.6,4.14,4.15,4.16,
4.19,4.20 etc. indicate the position of the spots, where excessive local heating
is likely to appear. This heating may be linked with the field H ms on the
surface of an element. After solving the approximate nonlinear thermal
equilibrium equation (see Chapter 5), a graph can be produced (Fig. 4.21)
which helps to evaluate a permissible tangential value of Hms,perm for a given
permissible temperature tperm ([4.22], p.381). This temperature is normally
recommended by international or national Study Committees of CIGRE (e.g.
SC No 12 'Transformers').
Having calculated the distribution of H ms one can easily identify potential
'hot spots' on the graph, where special precautions (temperature detectors,
Hor.=190, Vert. =75.000 Arrows-Turn.
(d)
jDATA
I
d24cfe5
flL1
!Max P=45
2
[kW/m 1
Min P=O.
P total
39.700
kW
TRANSFORMER 240 MVA - WITH FE SCREEN
Fig. 4.20 Interactive design of the tank screens with RNM-3D [4.22, 4.26j: (a) combined
copper (eu) and magnetic (Fe) screen, (b) H max for the case without any screen, (c) the best
variant for a frame screen, (d) loss density distribution and total stray losses in a five-limb
transformer with continuous Fe screen
non-magnetic steel, screening, directed cooling etc.), or more accurate

calculations should be undertaken. For instance, the three-dimensional
thermal analysis (see [4.22], p. 386-389) has shown that strongly non-uniform
stray loss distribution can increase the value of Hms,perm even by a factor of
two.
Reluctance Networks and Finite Elements 171
[~ni]
200
180
140
100
80
60
40
20
o 50 100 150 200 t

perm Lc';j
fo
Fig. 4.21 Permissible values of the tangential component of Hms,perm on the surface of a
solid steel plate versus the permissible temperature tperm of the plate [4.22}
4.6 TESTING RNM ON OTHER ELECTROMAGNETIC

DEVICES
In addition to the experimental verification presented in papers [4.20, 4.22,
4.26] etc., further testing of the RNM approach on some other electromagnetic
devices has been carried out (Figs 4.22, 4.23). Most internationally recognized
benchmark problems for testing various 3D codes are collected and regularly
reviewed by the TEAM (Testing Electromagnetic Analysis Methods) Work-
shop. Nearly all TEAM Workshop Problems have been solved using the RNM
at the Technical University of Lodz. One of them is Problem No 7 (Fig. 4.22)
solved by RNM [4.28] in reasonable computing time on a PC, with an
extremely good agreement with experiments and other computation methods,
such as the finite-element nodal A -, nodal T-0. and edge A -, tested among
others by Kamerai, Fujiwara and Olszewski (Okayama University).
An excellent convergence and agreement with experiment (see Fig. 4.23) of
a quasi-3D calculation using RNM of a flux density, coil inductance and
tractive force of a linear reluctance self-oscillating motor, has been demon-
strated in reference [4.29]. The work was done at Sydney University in
collaboration with the Technical University of Lodz.
4.7 RELUCTANCE NETWORKS AND FINITE ELEMENTS

The reluctance network formulation may be considered equivalent to some
other methods, such as finite differences or finite elements, in a similar way
(a) M~sh dimensions
+ H- "
2~ II! 3! 38 32 .c) so so 2S 2S
F-T+=-Ff=-I I i
=:
(b)
Ci:C C 1_1_ L r--l- ..
r l i I" 1- 1- I" h= :
1-1 +- 1-
--- !-j ..
B,=~~==~= ~=
..
rLLlf---'------'~-fL4-_H 30 '0
_ _ 1 __ ..1 _
C::-').
~ 't" 3e
1 I
- -1- - , -
_ _ 1_ _ ..1_
, 1
..
~ ==,==l ==
'.0
(c)
•.< .-A,7rn~
I
f
..• .,. -
•.> T-
0.' --+ - -
.1" __ If _ ~__
0.0
-0.'
I I 1
-0..
" - - r - -.- - , I
-1.% "T - - r - -1- - .,. - - I
-1.6 -+ - - I- - -1- - -+ - I
-,.0 ..1 __ L __1__ ..1 __ I
Fig. 4.22 Solution by RNM-3D of the TEAM Workshop Problem 7 [4.28}: (a) the RNM
3-D mesh of a square coil over aluminium plate with a hole, (b) calculated normal flux
density along the i, j, ..., m axis, (c) eddy current density along the x-axis in the plate
to circuit laws being considered equivalent to field theory if infinitesimal

spatial increments are assumed. For example, the definitions of div, curl
and boundary conditions may be obtained in this way by applying Kirchhoff's
laws to magnetic circuits.
Consider a scalar Poisson's equation in two dimensions at constant
permeability
a2A a2A
ax2 + ay 2 = -Ill (4.39)
and a corresponding energy function
(4.40)
Reluctance Networks and Finite Elements 173
te foree (Il]
t6
000-00 mcoauteme"t,
~. ANu'model
o
n
o ~--,-,......,......-,.........,.......--,-,......,......-,.........~:..,
o 20 <0 60 eo 100 120 1<0 160 180
dlsplocemenl 01 the bar centre (mm)
(a) (b)
Fig. 4.23 A quasi 3-D calculation for a linear reluctance motor by RNM [4.29]:
(a) cross-section, (b) calculated and measured propulsion force versus bar position
This functional represents the energy (or the Lagrangian) of a system, which
in the finite-element formulation must be a minimum for variation of each
of the node potentials, of which a typical one is A k • If first-order elements
are assumed, the potential at an interior point of one is given by linear
interpolation between the three vertex values A j (i = 1, 2, 3)
1 3
A (x,y) = 2d j~1 (a; + b;x + c;y)A; (4.41)
in which the coefficients depend on the vertex coordinates, and d is the area
of the triangle (see Chapter 3). As the first-order description demands a linear
variation, a similar approximation will also apply to J, i.e.
1 3
J(x,y) = 2d j~1 (aj + bjx + Cjy)Jj (4.42)
The variation in energy when any typical node potential A k is changed must
be zero, which gives, for each value of k, an equation of the form [4.4]
fJ[
v
:lA
k
= be b (aekjA j -
I
f3ek;Jj) =0 (4.43)
where e is the element number, while i denotes the vertices and ranges over
three values in each element.
Consider now a linear magnetic network in which a typical component
carries a flux cf) and requires an mmf of F. The energy input required is
Wm = cf)FI2 = cf)2R I1 I2. If the network is replaced by its dual, cf) becomes a
potential quantity, and thus the stored energy in any component, in terms of
the potential values cf). and cf)2 at the two ends, is
(4.44)
I=Jt
2
(a) (b)
Fig. 4.24 Elements of a reluctance network equivalent to finite elements {4.4j: (a) linkage
of a reluctance network with a flow I, (b) network model with current and reluctance
linking
Varying a single node potential, say <1», therefore changes the stored energy
linearly
(4.45)
Since the magnetic vector potential A is the flux linkage per unit length,
comparing the last result with eqn (4.43) shows that the a coefficients
correspond to branch reluctances in an equivalent magnetic network
representing unit length (see Fig. 4.24). Moreover, this network stores energy
in exactly the manner required by the a terms in eqn (4.43).
4.8 TIME-DEPENDENT FIELDS

Ohm's law for magnetic circuits VIl = <l>R Il can only be applied to static or
quasi-static magnetic fields at a constant frequency f. For such fields magnetic
flux becomes a full analogue of electric current. In the case of transient fields,
however, it is better to associate the magnetic flux = JB' ds with the electric
flux'll = JD' ds rather than with the electric current i = P' ds [4.3; 4.12; 4.19].
The electric flux corresponds to an electric charge Q Cu 'II. Moreover,= =
i = dQldt = C duldt and therefore the analogue of electric current i is not the
magnetic flux but its rate of change in time. This is consistent with the idea
of a magnetic current JIl put forward by Heaviside
1.
11
= d<l>
dt
= C11 dV
dt
Il
or III = jro<l> = jroCIl VIl (4.46)
where CIl is the so-called magnetic capacitance and VIl is the magnetic
voltage. Hence Kirchhofrs magnetic law for a magnetomotive force F reads
(4.47)
Time-Dependent Fields 175
where
1
R=-
11 C11
is the reluctance
y y
L =-=-
11 J11 _d4>
dt
(a) y
x
0
R
.. R .!..
F
+- V--
R R R
R F
R .!..
+- V- c-
R F R R
R R F
;..
0 t)-f-
(b) 0.05
• • • •• measurements
6 - seg. coil model
0.04 2 - seg. coil model
0.03
0.02
0.01
0.00 ~_.,er"-
-ll.01 +._ _ """" ..,.

-240 -120 0 120 240
distance from the coil centre [mm]
Fig. 4.25 Experimental verification of flux density calculated using RNM along the axis of
a motor of Fig. 4.23 [4.29J: (a) model, (b) distribution in the gap without an iron bar
is the transference, which is a measure of the eddy currents induced by the

flux in all conducting paths which are wholly or partially linked with the flux,
and
N
Y= ~ ir
r=1
is the current linkage. The reluctance is treated as the inverse of the 'magnetic
capacitance' rather than a 'magnetic resistance'. Carpenter [4.3J concludes that
two electric-magnetic analogies are possible:
electric current: J", = d<l>/dt

electric voltage: mmfF mmfF
· L -di d<l>
u = R1+ F=R", +L"'dt
dt
u = Ri + [bJr i dt
y
resistance R R", L", =-
J",
Some interesting practical applications of the reluctance network method
for analysis of time-dependent two-dimensional fields are presented in papers
by Oberred [4.18J and Demenko [4.8J.
4.9 EXPERIMENTAL VERIFICATION OF THE METHOD

Experimental verifications of the reluctance network method and the
RNM-3D package have been reported on several occasions and by many
0.72
0.10 l(M)
0.' JOOnOI. _ •• "'.fft.... b
0." L(K] _ •• c•• _ ...'
........- 1 , _4,1
. . . . . . . . . . . . .IfII (0-$0-
0.66 0.'
0."
0",
0.&2
0.&0 0.'
O.!>I
0.)
O~
0.2
O~ +-...........~-.- ........~-.-......,.~....--......., O.t_:l:-'OO:--'"---C_Ir:-:OO-~O--.:TOO~~200

o :.0 100 ,:.0 200 2)0 3CO 3:.0
4ltFa.cc..cat ., t,J\....... ccAt-:- (ft\II\)
....I'nol cI'ltone! (EO) [mm)
(a) (b)
Fig. 4.26 Experimental verification of calculations by RNM of a coil inductance in a linear

motor of Figs 4.23 and 4.25 [4.29J: (a) convergence versus dimensions of the external open
boundary region, (b) calculated and measured inductance
Experimental Verification of the Method 177
II
(a) (b)
X
X I
II I
" I
I I 1l.
I Il
1)
e ~
I :z
I <
f-o
Lr.J
0::
0
U
I 4 r
I I
l I
I
I
I(llt •lilt .olr
o
Bw 0
B..... Br...
(c)
Fig. 4.27 Measured (dashed) and calculated (bold) fields in a 315 kV A transformer [4.26J:
(a) flux density Bn in L V and HV winding, (b) Bn on the tank wall (y = 0), (c) Bn along the
tank circumference
authors [4.2; 4.7; 4.20; 4.26; 4.27; 4.28; 4.29J. For example, an axisymmetric
problem is relatively easy to investigate (Figs 4.23b, 4.25 and 4.26) and has
shown excellent agreement with measurements. Even for an open boundary
problem (Fig.4.25) convergence has been fast (Fig.4.26b) and computed
results compared very favourably with experiment (Figs 4.23b, 4.25b and
4.26b). For a much more complicated TEAM Workshop problem 7
H (x=O,y,z=O)
msx
Fig. 4.28 Axial magnetic field distribution (HmJ along the tank circumference (x = 0) of a
150 kV A model transformer {4.26j
- - calculation using RNM-3D
x x x experiment
(Fig. 4.22), with 3D solenoidal field and eddy currents, results from the
reluctance network computations have been remarkably close to measure-
ments and calculations, using other methods, conducted at Okayama Univer-
sity, Japan [4.28].
The most successful and important application of the method so far, from
the engineering point of view, has been to large power transformers, as already
demonstrated in th~s Chapter. Relevant tests on real transformers, especially
measurements of power loss associated with leakage fields, are very difficult.
Nevertheless many results have been published and may be found for example
in references [4.22; 4.24]. Comparisons of calculations performed using
RNM-3D with such measurements consistently show good agreement. Some
further examples of such comparison are shown in Figs 4.27 and 4.28.
In conclusion it should be emphasized that although the reluctance network
method is overshadowed by other techniques, especially the finite element
method with its recent significant advances (see Chapter 7), it continues to
play an important role in engineering applications. The results produced by
this method are approximate, but they help the designer focus his attention
on critical issues and not be burdened with those which, from the practical
point of view, are less important. Moreover, the expertise of the designer and
his knowledge of the device, including existing simplified empirical or
analytical partial solutions, may be easily incorporated into the reluctance
network formulation. Most significantly, though, the method provides the
solution in extremely short computing times which makes it very suitable for
interactive design.
5
Field Synthesis
Krystyn Pawluk
5.1 INTRODUCTION
5.1.1 What is field synthesis?
The term field synthesis is introduced as the converse of field analysis. Field
analysis is usually easier and thus preferred, and can be applied when the
geometry and material structure of a device are known and when the field
sources are given. The analysis is performed by solving a field equation which,
generally, is a partial differential equation of the second order with respect
to a field quantity of a scalar or vector type. The solution of the analysis
problem refers to a region that is geometrically determined, i.e. the shape of
the region is given, and moreover, the region has its material characteristics
specified. It is assumed that the analysed field is generated by primary field
sources, which are given as boundary conditions or internal excitations.
The field synthesis is an inverse problem to that specified by analysis and
may be formulated in a variety of ways. The distribution of a field quantity
is assumed as given in the whole or part of the region under consideration and
we may be looking for:
• outer field sources defined as suitable boundary conditions;

• inner field sources, for instant coils or electric charges;
• material parameters or characteristics;
• the geometric shape of the region.
Thus, the synthesis problem consists of determining unknown effective

causes that influence the physical field, the distribution of which is either
postulated or previously obtained as the result of a measurement. In electrical
engineering we are generally interested in electrostatic, magnetostatic, elec-
tromagnetic or current-flow fields, but in addition, the synthesis of a thermal
field may also be the object of the design.
The above conclusions are summarized below:
(1) If synthesis is applied to the design of some part of an electrical equip-
ment (electrical machine, apparatus, isolator, antenna etc.) and the
desired properties of the equipment can be expressed as a function of
the field distribution in part of this equipment, then we shall call this
the design synthesis.
180 Field Synthesis
(2) If synthesis is performed to identify the geometry or material charac-

teristics of the existing equipment and this identification is based either
on the measured distribution of some field quantity on part of this
equipment, or on the measured integral quantity dependent on the
distribution of a field quantity (e.g. a magnetic flux), then we shall call
this procedure the identification synthesis. There is a very practical
aspect to such problems when, for example, we attempt to assess the
quality and accuracy of equipment after a long operating time or
following a fault.
(3) In the case when synthesis is performed in order to determine an
optimal feature of the equipment, we have the optimal synthesis.
The design and identification synthesis problems differ significantly in

their aims; nevertheless the mathematical formulations are often similar and
sometimes identical.
A common property of synthesis problems is that they are not unique. It
is generally because the information available is not sufficient to describe the
problem mathematically. In some cases the amount of input data may appear
to be sufficient, while in fact some equations are either linearly dependent or
even contradictory. In the case of deficient data we have an under-specified
problem, and for an excessive number of independent information we are
dealing with an over-specified problem.
The design synthesis problems may yield solutions which are technically
impossible. This may happen when a designer has insufficient experience in the
matter and improper input data has been chosen. The identification synthesis
problems, as they refer to existing objects, should always have a solution, but
even here the inaccuracies of measurements may lead to inconsistent solutions.
Most tasks concerning the design of electrical equipment are essentially syn-
thesis problems. A traditional approach avoids the proper synthesis formula-
tion altogether. Thus a conventional trial-and-error strategy consists of a
series of solutions of analysis problems complemented by empirical or semi-
analytical results and often verified by building and testing physical models.
An alternative new approach to design uses CAD and optimization, as
discussed in Chapter 7. In this Chapter, however, we concentrate on direct
methods of synthesis.
The same physical equations describe both the synthesis and the analysis of
electromagnetic fields, although the manner in which the problem is for-
mulated is different. Thus, the synthesis problems can be classified with
reference to the field problem under consideration:
(1) inverse problems to Laplace's equations, where no internal sources are

present (no electrical charges, no currents);
(2) inverse problems to Poisson's equations, where internal field sources
exist;
(3) inverse problems to Helmholtz equations, where current is induced in
the region;
Introduction 181
Fig. 5.1 A field synthesis region
(4) inverse problems to wave equations, where a full set of Maxwell's equa-
tions must be taken into consideration.
The following topological properties of the synthesis problem should be

noticed. In the region 0 under consideration, a subregion 0 0 C 0 is iden-
tified which we shall call the synthesis controlled subdomain, see Fig. 5.1. In
this subregion the field is subject to verification during the solution. The
domain 0 0 can be of the same or smaller dimension than 0, and can also be
split into a number of discounted parts 0 01 , 0 02 , ... , 00n.
Synthesis problems can be classified as falling into two categories:
(1) a group of rigid structure, when the geometry and environmental

characteristics of 0 are fixed and when the field in 0 0 is to be syn-
thesized by determining the field sources, Le. the boundary conditions
on r or/and the internal source distribution in the subregion Os,
called the internal sources subdomain;
(2) a group of supple structure, when the shape and environmental
characteristics of 0 are tentatively assumed, and the field in 0 0 is
then synthesized by final determination of its geometry and/or the
characteristics. Care has to be taken not to affect the uniqueness of
boundary conditions and internal source distributions.
When a problem is formulated in such a manner that it requires specifica-

tion of a single quantity or condition (e.g. boundary conditions only) then we
have a simple synthesis problem, whereas if more that one values are needed
it leads to a compound synthesis problem.
5.1.2 Development of methods and techniques

Synthesis problems were first solved using the least-square method due to
Gauss and Legendre. The modern view is that this method is not satisfactory
because it often leads to unstable solutions. The development of theory of
improperly- and ill-posed problems has been stimulated by progress in
numerical techniques. It is generally required that solutions of synthesis
problems should be stable and relevant to the physical nature of the objects
under consideration. Contemporary synthesis techniques follow the works of
182 Field Synthesis
Twomey [5.69], Phillips [5.42] and Tichonov and Arsenin [5.68]. These techni-
ques are known as regularization methods and lead to very effective computer
codes for problems described by integral equations, as well as those which are
derived from the finite element formulation.
It is very likely that methods for solving inverse problems of electro-
magnetism will continue developing as an independent field. It appears that
the majority of necessary theoretical work has already been done and effort
is now directed to creating appropriate computer programs and their imple-
mentation to the design of electrical equipment. It may be appropriate to point
out that even a simple calculation of the magnetizing current in machines and
transformers is a synthesis problem, Le. finding the localized magnetic field
source for a prescribed magnetic flux. The same aspect of synthesis may also
occur in analysis problems examined by the method of secondary sources (see
Pawluk [5.40)). This method uses the principle that the continuity of vector
field quantities on the internal and external boundaries is equivalent to a
suitable distribution of potentials (single or double layer) on these boundaries.
A sensitivity approach seems to be promising for a large class of design
problems - see Henneberger et al. [5.16], II-han Park et al. [5.17] and Sikora
[5.60]. A general treatment of the synthesis problems may be seen in the
papers by Guarnieri etal. [5.13], Salon and Istfan [5.55] and Simkin and
Trowbridge [5.66].
5.2 A SURVEY OF INTEGRAL EQUATIONS

5.2.1 Types of equations
In this Section we shall discuss such aspects of integral equations which are
relevant to the theory of synthesis of electromagnetic fields. Moreover, we
shall present only the simplest formulation for one independent variable,
although the treatment can be easily broadened to cover 2D and 3D field
problems. An integral equation is an equation in which the unknown function
appears under the integral sign. Thus the integration is performed on the
function.
The integral equation in which at least one limit of integration is a variable
x
JK(x, ~)g(~) d~ = u(x) (5.1)

a
is called Volterra's equation of the first kind. In the above equation g is the
unknown function, and the function U on the right-hand side is prescribed.
If u(x) == 0, eqn (5.1) becomes homogeneous. The function K is called the
kernel.
The integral equation
x
g(x) + A. JK(x, ~)g(~) d~ = u(x) (5.2)

a
A Survey of Integral Equations 183
is called Volterra's equation of the second kind. The unknown function g

appears twice in this equation: alone and under the integral sign. In the
equation A. is a constant. The particular cases of Volterra's equations are
found when limits of integration are fixed. This leads to
Fredholm's equation of the first kind:

b
JK(x, ~)g(~) d~ = U(x) (5.3)

a
Fredholm's equation of the second kind:

b
g(x) + A. JK(x, ~)g(~) d~ = u(x) (5.4)

a
The kernel of an integral equation may be classified as:
(l) symmetric, if
K(x,~) = K(~,x)
(2) regularized, if
bb
JJ IK(x,~) 12d~dx
aa
exists and is not zero,

(3) midcontinuous, if
b
lim J IK(x + t:.x,~) - K(x,~) 12d~ = 0

~ ....o
a
(4) degenerate, if it can be expressed as

K(x,~) =f(x) ·h(~)
or in the form of a sum of such products,

(5) if the kernel is complex K(x, ~) then it will always have its conjugate
counterpart K* (x, ~),
(6) the transposed kernel to K(x, ~) is defined as
KT(X,~) = K(~, x);
for the symmetric kernels we have
K T (x,~) = K(x, ~),
(7) if the complex kernel K(x, ~) has its conjugate K*(x, ~) equal to the
transposed kernel K(~, x), then it is called a Hermitian kernel. Clearly,
real and symmetric kernels are Hermitian,
184 Field Synthesis
(8) when the kernel is singular, i.e. when it has the form
then the relevant integral equation is said to be singular.
Integral eqns (5.1)-(5.4) are called linear equations. If the expression under
the integral sign cannot be written as a product of the kernel and the unknown
function, for instance
b
JK[x,~, g(~)] d~ = u(x) (5.5)

a
then such equation is said to be nonlinear. In the case of eqn (5.5) it is the
nonlinear Fredholm's equation of the first kind.
An integral equation can be written in a compact form by the use of an
integral operator:
b
K = JK(x,~) ... d~ (5.6)

a
Thus the Fredholm's equation of the first kind reads

Kg(x) = u(x) (5.7)
The operator K attributes a function f(x) (belonging to the functional space
9 of some class of the unknown functions) to a function u(x) (which belongs
to the space <u of given functions). This attribution or mapping may be
symbolically written as K: 9 -+ <u.
Eigenfunctions and eigenvalues

If for x e (a, b) a non-negative function h;(x) satisfies a homogeneous integral
equation of the second kind (for example the Fredholm's eqn (5.4» at any
fixed value of the parameter A;, so that
b
h;(x) + A; JK(x, ~)h;(~) d~ = 0, where i= 1,2, ... ,m (5.8)

a
then such functions h;(x) are called the eigenfunctions of the equation and
the relevant parameters A; are called the eigenvalues. The largest absolute
value of l/A; is termed the spectral radius of the operator K and it is written
e(K). Thus:
e (K) = I~J = max [ I~1 I'I~J '.. ,I~J J (5.9)
On the solution of Volterra's equations

It has been proved that if in the integral equation of the second kind (e.g.
eqn (5.2) or eqn (5.4», having the midcontinuous kernel K(x, ~), the following
condition holds:
(5.10)
then this equation has a unique and continuous solution gl (x) for an
arbitrary right-hand side function u(x); this solution can be obtained using
the method of successive approximations. The above theorem is particularly
important for the case of Volterra's equations of the second kind when
e(K) ~ 0, so that eqn (5.2) has a unique and continuous solution which can
be found by successive approximations.
Some Volterra's equations of the first kind (5.1) may be reduced, by
differentiation with respect to x, to the equivalent Volterra's equation of
the second kind, and hence their solutions may be found. In order to follow
this procedure we must have u(x)x=a = 0 and K(x, ~)~=x = K(x, x) =1= O. The
singular Volterra's equation does not fall into this category. The differentia-
tion of eqn (5.1) may be performed as follows
x
~ jK(x, ~)g(~) d~ = dU:)

a
rJ aK(x,~)
ax g(~) d~ + [K(x, ~)g(~)h=x - [K(x, ~)g(~)h=a =
dU(x)
<LX
a
After appropriate substitutions and division by K(x, x) we have

x
j
g(x) + K'(x, ~)g(~) d~ = v(x) (5.11)
a
where the new kernel is
K' (x,~) = I aK(x, ~)

K(x,x) ax
and the right-hand side function becomes
_ dU(X)]
v(x) = K(x,1 x) [
K(x, a)g(a) - <LX
Solution of Fredholm's equations
The following properties of the Fredholm's equations of the second kind (5.4)
determine the question of their stability:
(I) if A. is not the eigenvalue of eqn (5.4), then the equation has a unique
solution h(x) for an arbitrary function u(x);
(2) if A. is the eigenvalue of eqn (5.4) then the unique solution exists only
for certain functions u(x) and only for Hermitian kernels.
186 Field Synthesis
The solution of the Fredholm's equation of the second kind may be obtained
using many methods including:
(1) the method of successive approximations,

(2) the Hilbert-Schmidt method with the use the resolvent kernel,
(3) the Fredholm's method with the use the resolvent kernel,
(4) the kernel iteration.
Readers wishing to study the above methods are referred to the book by Korn
and Korn [5.21].
When inverse problems of electromagnetic fields are formulated, they
usually involve Fredholm's equations of the first kind. If these equations have
Hermitian kernels then they may be solved by expanding them into infinite
series in terms of their eigenfunctions h;(x). Generally, the solutions of
Fredholm's integral equations of the first kind occurring in electromagnetic
field synthesis are neither unique nor stable. In recent years many numerical
algorithms for the solution of such equations have been developed; the method
of regularization for example has been efficiently used in some applications,
see Section 5.4.
5.2.2 Green's functions

In the theory of integral equations of mathematical physics, Green s functions
I
play an important role. Potential theory and Green's functions are closely
related. A detailed study of the potential theory is beyond the scope of this
book. Nevertheless, in this Chapter we shall introduce some fundamental con-
cepts underlying the Green's functions. In Chapter 3 Green's functions have
already been used to describe the boundary element method. In this Chapter
we shall give some more consideration to Green's function encountered in
inverse electromagnetic problems.
Let us consider Laplace's equation
tJ.u =0 (5.12)
In the open n-dimensional space Q(n) this equation has the solution
(5.13)
which is a harmonic function, i.e. it is twice continuously differentiable,
and, if n ~ 3, it is regular in Q(n). This solution is known as a fundamental
solution, if it is presented in the form
(5.14)
where
(5.15)
is a modulus of the position vector, i.e. r is the distance between the following
two points:
(1) the first point P = P(XI, ... , x n ), or in an abbreviated notation

P = P(P), at which the field is to be determined, and
(2) the second point Q = Q(~J' ... , ~n) or Q = Q(Q), that is unique in
o(n), at which the fundamental solution does not obey the Laplace's
equation.
Thus, for n ~ 3, the fundamental solution has an inherent singularity at

P -+Q, Le.
(5.16)
The fundamental solution (5.14) belongs to a class of functions known as

Green's functions, which satisfy the Laplace equation. We could say that the
fundamental solution is the most elementary of the Green's functions.
We shall demonstrate now how to obtain in a simple way the explicit funda-
mental solution of Laplace's equation in 0(3). Writing Laplace's equation in
spherical coordinates r, e, n we have
2
-1- ar [-
2 auJ
+-- -1 sm a [ .e-
auJ+ 1 au
r 2 ar ar r 2 sin eae ae r 2 sin2 e -an 2 (5.17)
Because the solution has a singularity at the origin of the coordinate

system, let us express it as a function of the r coordinate only, so that
aulae = aulan = O. Thus the following ordinary differential equation in terms
of r is found
2
u"+-u'=O (5.18)
r
the solution of which is given by u(r) = C 1 + C2 1r, where C 1 and C2 are

arbitrary constants. In the fundamental equation we usually put C 1 = 0 in
order to preserve the regularity of the solution. We also require that the fun-
damental solution should satisfy Poisson's equation, where the source is
described by the unit impulse also called the Dirac delta function
dU= -8(P-Q) (5.19)
Thus, if we put C2 = 1/41t, we finally obtain

I
w3(r) =-4 (5.20)
1tr
so that the fundamental solution is identical to the Green's function of
Laplace's equation in 0(3). The integral of d W3 (r) over the whole 0(3) space
(or any sphere of radius R) is
J d W3 (r) d~ dT) d~ = -1 (5.21)

Q(l)
which is a consequence of the property of the Dirac's delta function. Instead,

let us introduce an additional function
188 Field Synthesis
(5.22)
and find its Laplacian
(5.23)
Finally, we integrate over the sphere of radius R and put e -+ 0, which

yields
21l 1l R 2
lim
&-0
J Llw3(r,e) dO = lim
&-0
J JJ - 4n
-J( r3~ + e2)5 r 2 sinedrdeda
01/> 0 00
= !~"J [- e2-J(~2R: e2)3] = -I (5.24)
We conclude that w3(r), as given by eqn (5.20), is indeed the fundamental

solution in 0(3) of the Laplace's eqn (5.12), except for the singular point
r = 0, or of the Poisson's eqn (5.19), too.
In the open 20 space 0(2) a similar discussion produces an ordinary dif-
ferential equation
I
u"+-u' =0 (5.25)
r
leading to the following 2D fundamental solution:
I I
w2(r) = -In- (5.26)
2n r
which, like W3 (r), has a singularity at r -+ O. The difference in relation to

w3(r) is because w2(r) is not regular for r -+ O.
A fundamental solution for the Helmholtz equation can be found in a
similar manner. For example, in the case of a mono-harmonic electromagnetic
wave in vacuum, the Helmholtz equation reads
Llu + J(2u =0 (5.27)
where k 2 = oh10Eo' The ordinary differential equation for u(r) in 0(3) is
found
2
u"+-u'+k 2u=0 (5.28)
r
J
Substitution of u(r) = r v(kr) leads to the Bessel's equation
v" + :r v' + [I - 4(~r)2] v = 0 (5.29)

which has a solution

C
v(kr) = Ji- J_ I12 (kr)
with the presence of the Bessel's function of the minus half order of the
first kind. Thus the fundamental solution of wave eqn (5.27) in 0(3) may be
found as
1
lf3k(r) = -4
1tr
exp( ±jkr) (5.30)
where the signs' +' or '-' refer to the forward or backward travelling waves
respectively. Hence the real fundamental solution may be taken as
cos (kr)
W3k ()
r = 2kr (5.31)
leading to the value W3k = 1I41t on a sphere with radius equal to the
wavelength r = A. = 21t/ k.
The solution of eqn (5.27) in 0(2) may be found in the form of Hankel's
function of zero order
(5.32)
where for the forward wave we use the minus sign and s = 1, and for the
backward wave choose the plus sign and s = 2. For the real fundamental
solution we obtain the Bessel's function of the second kind, following Weber's
definition w2k(r) = CYo(kr). Finally, let us consider the diffusion equation
flu - k 2 -
au = 0 (5.33)
at
where k 2 = ~a, describing the penetration of some scalar quantity into a con-
ducting body which is able to resist such penetration, e.g. in a conductor. The
fundamental solution in 0(3) is found as
2 22
k [ k rJ
vn(r, 't) = '\j(41t't)n exp - 4t (5.34)
where 't = T - t is the time interval between the instant of excitation and the
instant of observation.
If the monochromatic wave is subjected to the diffusion, the Helmholtz
equation is
!1u - jk 2 u = 0 (5.35)
where k = ro~a and the fundamental equation in 0(3) is
~3 ( r) = ~ [ber( kr) + j bei ( kr) ] (5.36)
and in 0(2)
~2(r) = C[ker(kr) + j kei(kr)] (5.37)

190 Field Synthesis
z,(
r g(~.1].()
...- P(x, y, z)
\Q(~'7J'\)
u(F)
o
I(Q)
x, ~
y,77
Fig. 5.2 Region n with the boundary r
Green's function in bounded regions

It is well known that the potential at point P (x, y, x) E n, as shown in Fig. 5.2,
where n is a 3D (or 2D) region bounded by the surface (or line) r, may be
expressed in terms of three quantities:
(1) the source function f(~, 11, ~), in n;

(2) the distribution of the normal derivative of the potential on the
boundary surface (boundary line) r,
(3) the distribution of the potential on r.
In relation to (1), the source function may for example be:
• a distribution of volume density of electric charge in electrostatic fields,

• a distribution of heat sources in temperature fields,
• a distribution of d.c. current density for static magnetic field problems.
With reference to cases (2) and (3), the normal derivative of potential and the
potential itself could be:
• the normal component En = -a vI an of the electric field strength E

owing to an electrostatic potential V in a 2D or 3D region,
• the normal component H n = -a Vm I an of the magnetic field strength H
owing to a magnetostatic potential Vm in a 2D or 3D region,
• the tangential component B( = aAzlan of the magnetic flux density B
owing to a magnetic vector potential A = Azk in a 2D region.
In general the potential u(x, y, z) at P (x, y, z) E n, which normally is written

in abbreviated notation as u(P), can be expressed as a superposition of three
potentials:
• the volume potential in 3D regions (or surface potential in 2D regions)

called the Newton's potential, dependent upon the field source distribu-
tion in 0,
• the double layer potential, dependent upon the prescribed distribution
of the normal derivative of the resulting potential on r,
• the single layer potential, dependent upon the prescribed distribution of
the potential on r.
The elementary volume potential du(P) in the point P (x, y, z) is given by
du(P) = j(Q) dQ (5.38)

41tr(P, Q)
where the following abbreviated notation is used
dQ= d~, d1'\, d~
r(P, Q) = \1'[ (x - ~)2 + (y - 1'\)2 + (z - ~)2]

For example in the case of an electrostatic field
dV(P) = o(Q) dQ
4m:r(P, Q)
where the source functionf(Q) is the volume density of electric charge divided
by 41t times the electric constant.
The elementary double layer and single layer potentials at P (x, y, z) are
given by
1 og(Q)
du(P) = 41tr(P, Q) an dQ (5.39)
and
du(P) =- :n [41tr(~,Q)]g(Q)dQ (5.40)
where g(Q) = g(~, 1'\, ~) is the given boundary potential at Q(~, 1'\, ~) E r. The
total potential will be determined by integration over 0 U r.
It can be seen that if in the unbounded 0(3) it is assumed that
lim f(Q)
Q-oo
=0
then the volume potential is the only potential and is given by integration over
the entire 0(3)
f(Q)
u(p) = J41tr(P, Q) dQ (5.41)
192 Field Synthesis
Q
Q)
e
~I
til.
'01
til·
.~ I
Q' Q
r'(p,Q' .
I
I ep
I
I
Fig. 5.3 Image of the field source point
The function of two points (and thus of six variables) of the form
1/4nr(P, Q) can be recognized as the kernel in the integral transformation of
eqn (5.41). This kernel is singular because
lim r(P, Q) = O.
Q-P
In the case of regions with closed boundaries, or boundaries limited by, say,
lines of symmetry, the kernel will be influenced by the appropriate boundary
conditions or types of field symmetry. Let us study a simple case of a sym-
metrical field in 0(3). The field symmetry axis is shown in Fig. 5.3, forming
the half-space O~). As a consequence of field symmetry we can determine a
point Q', which forms an image of a point Q relative to the symmetry axis,
where we havef(Q') =f(Q). The integration can be performed either over the
entire space 0(3) with the use of the kernel 1/4nr(P, Q), or can be limited to
0~3), with the kernel
1[1
2n r(P, Q)
1]
+ r(P, Q')
Both the above kernels can be interpreted as Green's functions.
Similar considerations will apply to the choice of Green's functions in the
electrostatic or magnetostatic fields in regions limited by planes separating
two media of finite and infinite (idealized) electric or magnetic constants.
The Green's functions are also affected by the boundary conditions (Dirichlet,
Neumann, mixed).
Generally, we notice that the total potential at P(x,Y, z) can be expressed
as the sum of three components: the volume potential, the double layer poten-
tial and the single layer potential. We have, for each component part, a kernel
composed of the fundamental solution 1/41tr(P, Q) and the functional supple-

ment F(P, Q), which depends upon the shape, material characteristics and
boundary conditions inherent to the problem under consideration. Such a
kernel can be interpreted as a Green's function, and for 3D problems it is
usually written as
I
G(p, Q) = 41tr(P, Q) + F(P, Q) (5.42)
In order to determine the main property of Green's functions, we notice that

the Green's function expresses the potential at P under the assumption that
the source function tends to unity at a unique point QI and disappears at all
other points Q
f(Q) = {1, =
for Q QI (5.43)
0, for Q"* QI
This can be described with the use of the Dirac delta function o(P, Q) =
o(x - ~)o(y - 11)O(z -~) as
J o(P-Q)dQ= {~' for QeQ

for QeQ
(5.44)
o '
so that
~G(P, Q) = -o(P - Q) (5.45)
where the Laplacian is expressed in terms of variables ~, 11, ~. Expression (5.45)
simply states that the Green's function satisfies Laplace's equation almost
everywhere, except the singular point at which it obeys Poisson's equation with
unity source function.
The second important property of Green's function is that
G(p, Q) > 0 in Q (5.46)
and the third property is its symmetry:
G(p, Q) = G(Q,p) (5.47)
It can be verified by inspection that the fundamental equation 1I41tr(P, Q)
has all three properties mentioned above. As a consequence, the function
F(P, Q) occurring in eqn (5.42), will also possess these properties.
General equation for potential

Following the considerations of the previous section we can write the potential
at P(x,y, z) e Q with the boundary r (see Fig. 5.2) in the following form
u(P) = 1 G(P,Q)f(Q)dQ +! [G(P,Q) ag~;) -g(Q) aG:;';Q)]ds

(5.48)
where:
G(P,Q) Green's function,

g(Q) boundary potential,
194 Field Synthesis
n, s denote the variable normal to f, and the surface (or line)

boundary element, respectively.
Making use of the fact that there is a certain degree of arbitrariness in the
choice of the Green's function, it is possible to introduce some special condi-
tions for G(P, Q) leading to such a form of the potential formula which will
contain the dependence upon just one type of boundary condition - the one
that actually occurrs in the problem.
For Dirichlet boundary conditions (i.e. for the given potential g(Q) on f),
Green's function could be selected to satisfy:
G(p, Q) = 0, P,Qef (5.49)
and then the first term under the sign of integration will disappear.
For a Neumann boundary condition (i.e. for the given og/on on f) a similar
requirement may be introduced
oG(P, Q) =0 P,Qef; P*Q, (5.50)

on '
However, in the case where f is a closed surface (or a closed line), the non-
uniqueness of the Neumann boundary problem necessitates supplementary
specifications of the average potential on f.
For the Dirichlet boundary problem the surface Green 's function may be
defined as
.
Gs(P, Q) = _ aG(p, Q)
an
IQer
(5.51)
where n is an external normal to f.

Let us express Gs(p, Q) for two types of bounded regions: a half-space O~)
and a half plane 0~2). According to eqn (5.20) we have the following G(P, Q)
as the fundamental solution in 0(3)
1
G(P, Q) = 41tvf[ (x _ ~)2 + (y _ 1'\)2 + (z _ ~)2] (5.52)
In the case of half-space z > 0, i.e. when f is a plane z(x, y) = 0, the external
normal to f is -z. Thus, following the idea of an image we have
It can be easily verified that for z -+ 0 we have G(P, Q) -+ 0, so that the

condition of eqn (5.43) holds. It can also be shown, by differentiating twice
with respect to the variables ~,1'\,~, that the function G(P, Q) satisfies
Laplace's equation.
°
Following the definition of eqn (5.51) let us now formulate the surface
Green's function for half-space z >
Gs(p, Q) = _ iJG(P, Q) = iJG(P, Q) I

iJ ( -~) iJZ ~=o
z (5.54)
Similarly we can find Green's function for a plane

1 1
G(p, Q) = 21t In .J[ (x _ ~)2 + (y _ 1'\)2] (5.55)
In the case of a half-plane y > 0, where r is a line y(x) = 0, we have
G(p, Q) = 21t
1{~)2 +I1'\)2]
In .J[(x _
I~)2}
+ + 1'\)2]
(y _ -In .J[(x _ (y
(5.56)
or
(5.56a)
The surface Green's function becomes
G (p Q)
s ,
= _ iJG(P, Q) = iJG(P, Q)
0(-1'\) 01'\
I n=O
= y
1t[(X_~)2+y2]
(5.57)
Finally let us examine the Green's function for an inside of a cylinder of
radius R. Starting from the Green's function in Q(J), we may express eqn
(5.52) in cylindrical coordinates, then for the points P(r, n, z) and Q(o, ~, ~)
we obtain, after relevant substitutions,
1 1
G(p, Q) = 41t .J[r2 + 02 _ 2rocos(n _~) + (z _ ~)2] (5.58)
The Green's function for the region limited by r < R can be found by taking
into consideration the images on the inner surface of the cylinder
1 {
G(P,Q) = - 1 -
41t :v'[r 2 + 0 2 - 2rOcos (n -~) + (z _ ~)2]
:.)[r 2 + (2R - 0)2 - 2r(2R ~ 0) cos (n - ~) + (z _ ~)2]}

(5.59)
The surface Green's function for the inner cylinder surface may be obtained
by differentiating eqn (5.58) with respect to the external normal 0 and then
substituting 0 = R
196 Field Synthesis
Gs(P, Q) = _ aG(p, Q) I
at} Q=R
1 R - r cos (a - 13)
=- 41t ..J[r2 + R2 + 2rRcos(a -13) + (z _ S)2]3 (5.60)
When the point P is situated on the axis of the cylinder (i.e. the z axis) then
r = 0 and the function Gs(p, Q) has the form
1 R
Gs(P, Q) = 41t ..J[R2 + (z - S)2]3 (5.61)
5.3 TYPES OF PROBLEMS

We use the notion of field synthesis, as the opposite to field analysis which
is commonly considered as a mathematically modelled physical problem
that consists of solving the field equation, which, generally, is a partial dif-
ferential equation of the second order with respect to some scalar or vector
field quantity. In analysis this solution is accomplished in the environmentally
and geometrically specified region which may be finite, bounded infinite, or
unbounded infinite. Field sources in the analysis are known and given in the
form of boundary conditions or/and internal sources.
The synthesis of the field can be formulated in many ways and we must
define precisely what kind of problem is to be investigated. The field distribu-
tion, or some integral of it, is given a priori in any synthesis problems. Some
or all of the following quantities synthesizing the field may therefore be
investigated:
(1) boundary conditions (Le. outer field sources),

(2) inner excitations (Le. inner field sources),
(3) geometric shape of a region including the localization of inner sources,
(4) material parameters.
5.3.1 Inverse problem to Laplace's equation

In this section the basic formulation of a synthesis problem is presented as
the inverse to the problem described by Laplace's differential equation with
n
Dirichlet boundary conditions. A closed 3D (or 2D) region with a boundary
surface (or boundary line) r is shown in Fig. 5.4. Let r be the surface of a
fixed shape, hence the considered problem belongs to the group of a rigid
structure. Suppose that in n a scalar potential field u(P) may exist satisfying
Laplace's equation
~u(P) = 0 for Pen (5.62)
Let the following considerations be limited to the functions u e CU, where
CU is a metric space with a distance function t}'\l[u(P), u(P)] defined as a
metric. Suppose that in the subregion no c n, called the synthesis controlled
Types of Problems 197
z,(
g(Q)
r
{1
x,~
y,Tj
Fig. 5.4 Region in which Laplace's equation holds
subdomain, the function ii(P) is prescribed. The synthesis problem, which we

wish to consider here, consists of searching for such a boundary function
g(P) = u(P)Per, that the solution u(P) satisfying Laplace's equation (5.62) and
uniquely dependent on g(P), would be in no as close to ii(P) as possible. This
condition implies the metric 'I'll = min for P e 00. It is evident, that the
solution of the problem is not unique because it depends on the choice of Q'U •
If the identity u(P) == ii(P) was required for the solution, then the choice of
ii(P) would be limited to functions satisfying accurately the Laplace's equation
with the given Dirichlet boundary conditions, i.e. to the unique harmonic
functions only. In practice, various functions ii(P) are taken into consideration
and accurate solutions are rarely sought.
Suppose that u(P) represents a scalar potential function. Then the formula
(5.48) can be applied with the first term becoming zero. The second term
can be reduced to zero too, if G(P, Q) is chosen in such a way that
G(P, Q)Qer = 0, see eqn (5.49). In consequence, after the appropriate choice
of Green's function, ii(P) will be linked with g(P) via a simple integral formula.
For the inverse problem, i.e. when ii(P) is a given function and the Dirichlet
boundary conditions g(P) is unknown, this integral formula becomes the
integral equation, or more specifically the Fredholm's equation of the first
kind similar to eqn (5.3), but formulated for a 3D (or 2D) region. Thus, the
unknown function g(P) satisfies the Fredholm's equation of the first kind
JK(P, Q) g(Q) dQ = ii(P), PeOo (5.63)

r
where the kernel K(P, Q) is a function of two points: the field point P e no
and the source point Q e r. It is necessary to differentiate between the
coordinate symbols of these points, namely P(x, y, z) and Q(~, 11, ~). Hence,
eqn (5.63), valid in a 3D subregion no, is
198 Field Synthesis
JK(x,y,z,~,T],~)g(~,T],~) d~dT]d~ = u(x,y,z) (5.64)

r
where x, y, z are the coordinate symbols in n as well as in no, and ~, T], ~ are
the coordinates of the same absolute coordinate system describing the point
on f. When a local coordinate system s, t is introduced on f, the integral
equation takes the form
J K(x,y,Z,s, t) g(s, t) dsdt = u(x,y,z) (5.65)

r
and is valid together with the transformation equations of the coordinates.
Similarly, we have for 2D regions
J K(x,y,~, T]) g(~, T]) d~dT] = u(x,y) (5.66)

r
and with the use of a local coordinate s on f
J K(x,y, s) g(s) ds = u(x,y) (5.67)

r
A kernel in the Fredholm's equation must be a known function. If both the
geometrical structure and physical features of n were given, one would be
able, in principle, to determine the kernel. In practice, it is possible only for
simpler shapes of n, as the explicit form of kernel for a general case is difficult
to obtain. The kernel occurring in eqn (5.63) is the Green's function of
the second kind K(P, Q) = Gs(p, Q) (see Section 5.2.2) called the surface
Green 's function, which satisfies the Laplace's equation
!J.Gs(p, Q) = 0, Pen, Qef (5.68)
with the homogeneous Dirichlet boundary condition
Gs(P, Q) = 0, P,Qef, P*Q (5.69)
and the following unique kernel integral equation:
JGs(p, Q) dQ = -1, P,Qef (5.70)

r
A Fredholm's equation is said to be linear (see Chapter 5.1.1) if the kernel
is independent on the unknown function g. In consequence, the synthesis
problem under consideration can be formulated by the use of the linear
Fredholm's equation when:
(1) the problem belongs to the group of rigid structure, and

(2) material parameters of the region n are independent of both the field
function u and its gradient.
Other problems
We will now specify some variants of the synthesis problems, all being inverse
to the Laplace's equation, and all having a mathematical formulation not

much different from that already described.
(I) When the dimension of no is less then that of n, the reduced number
of coordinates occurs in the kernel. This is the case in numerous
engineering problems. The simplest problem is when n and no are
assumed as 2D and ID regions respectively and a single boundary
coordinate s is determined on r. Then the synthesis problem is simply
formulated by the one-dimensional Fredholm's equation of the first
kind, like eqn (5.3), namely:
JK{x,s} g{s} ds = u{x}, xeno. (5.71)

r
(2) There are in n some number of split synthesis controlled subdomains

no;, where i = 1,2, ... , n.
(3) The unknown function of r is of a different type from the Dirichlet
boundary condition g(P) = u(P)Per. In the case of the Neumann
boundary condition the kernel would be a Green's function, Le.
K(P, Q) = G(P, Q).
(4) The function prescribed in no is not a potential u but its gradient, for
instance V(P) = -grad (P).
(5) The vector field U(P) , satisfying the vector Laplace's equation
!J.U(P) = 0, occurs in n and the distribution U(P) is prescribed in no.
(6) A case like the one mentioned above but with W(P) = curl U(P)
prescribed in 0 0 •
(7) n is not a closed region but it is a limited space (for instant the half-
space or half-plane). r is in that case the limiting surface (or line).
5.3.2 Inverse problem to Poisson's equation

In this Section we shall consider a basic formulation of the synthesis problem
that is inverse to Poisson's equation with homogeneous Dirichlet boundary
conditions.
Suppose that in a closed region n, with boundary r shown in Fig. 5.5,
a scalar potential field may exist satisfying Poisson's differential equation
=f{P}, Pen
!J.u {p} (5.72)
In a synthesis controlled subdomain no c 0 let the function u(P) be
prescribed for P e no. The synthesis problem is to determine such a source
function f(P) in n =
that the function ii(P) u(P)PeO' which is a unique
solution of Poisson's equation (5.72) subject to the homogeneous boundary
conditions g(P) = u(P)Per = 0, would be in no as close to u(P) as possible,
in the sense of the metric chosen for the problem
'I'U[ii{P}, u{P}] = min, PeOo (5.73)
We note that the integral formulation considered above is generally covered
200 Field Synthesis
z,(
r
-+ P(x,y,z)
!1u(P)= 0
-+Q( ~,1),()
~
\J};) I(Q)
x,~
y,1)
Fig. 5.5 Region in which Poisson's equation holds
byeqn (5.48), but only the first term needs to be applied. Thus the unknown
source function f(P) satisfies Fredholm's equation of the first kind
JK(P, Q)f(Q) dQ = u(p) PeOo (5.74)

o
where the kernel K(P, Q) is a function of two points: the field point P e 0 0
and the source point Q E 0, and the kernel here is simply the Green's function
K(P, Q) = G(P, Q), which satisfies the Laplace's equation
dG(P, Q) = 0, P,QeO; P* Q (5.75)
with the homogeneous Dirichlet boundary conditions
G(P,Q) = 0, P,Qef; P* Q (5.76)
and the following unique kernel integral equation
J G(P,Q)dQ=-1 (5.77)
o
Fredholm's equation (5.74) is linear when the material parameters of 0 are
independent on both u(P) and grad (P).
Other problems
(1) A field function can be controlled in the whole 0, Le. 0 0 = O.
(2) Two subregions are given in 0: a synthesis controlled subdomain 0 0
and a source subdomain Os, in which the unknown source function
is to be determined (see Fig. 5.6). Then the source function in 0 - Os
is given and, in many practical cases, it is simply zero.
(3) There are in 0 a number of subdomains 0 0 and Os, which mayor
may not be separate from each other.
z,(
r
+P(x,y,z)
11
+Q(~,?'],()
11 f(Q)
z
x, ~
y,?']
Fig. 5.6 Region in which the field is synthesized by the internal sources
(4) Boundary conditions on r are of different type than the Dirichlet

boundary conditions.
(5) V(P) = -grad(P) is prescribed in 0 0 ,
(6) A vector field U(P) exists in 0 and U(P) or W(P) = curl U(P) are
prescribed in 0 0 ,
5.3.3 Variable boundary problem

In this Section we shall discuss a basic synthesis problem applied to the group
of supple structure that is inverse to the Laplace's boundary problem with
Dirichlet boundary conditions.
Let 0 shown in Fig. 5.7 be a closed region with a boundary that is separated
into two parts r = r J U r 2 • Suppose that in 0 may occur the scalar field
u(P) C CU satisfying a Laplace's equation and suppose also, that in the syn-
thesis controlled subdomain 0 0 C 0 the distribution of the field function
U(P) is prescribed. The following Dirichlet boundary conditions are given on r
u(p) = gl (p), Perl (5.78)

u(p) = g2(P) = g2 = const,
r J is assumed to be fixed, while r 2 is variable and the equation y(P) = 0
describing this surface is to be determined in the synthesis problem. A constant
function g(P) on r 2 with the Dirichlet boundary conditions (5.78) has been
chosen in order to conserve this boundary condition independently from y.
A set of the available y functions should be determined. Let us impose two
limits r -2 and r 2' on the shape of y. In practical cases these limits will be set
as evident requirements by a designer. Figure 5.7 shows the region under con-
sideration for which 0 0 is independent of the variable part of the boundary.
202 Field Synthesis
r.'
2
r =?
2 •
r'
2
Fig. 5.7 Forming the boundary shape
The synthesis problem is to determine such a boundary shape f 2 con-

forming to the imposed limits fi and fi', or to put it in a different way, to
determine such a function Y comprised between the given limits, that the
solution u(P), satisfying Laplace's eqn (5.62) subject to boundary conditions
(5.78), would be in f o as close to u(P) as possible in sense of the metric.
The Fredholm's equation valid for the above synthesis problem is
gd K 2[P,Q,Y2(Q)]dQ=u(P) -u1(P), PeQ o (5.79)

[2
where
Ul(P) = J K1(P,Q)gl(Q)dQ
[,
can be obtained by integration of two known functions K 1 and gl. The

characteristic property of Fredholm's eqn (5.79) is that the unknown function
Yl occurs under the kernel. Thus the variable boundary synthesis problem
leads to the nonlinear Fredholm's equation of the first kind.
y,?]
Fig. 5.8 Forming the boundary shape described by y = y(x)

Let us study the above problem, Le. Laplace's equation (5.62) and boundary
conditions (5.78), in a particular case for the 2D region shown in Fig. 5.8. We
suppose that the equation of the boundary line r 1 can be written in the
explicit form y = YI (x) and, similarly, the equation of Y2 can be of the form
y = Y2(X), with the supposition that Y2 C 'U E , the elements of which are half-
ellipses of the major axis fixed to a and of the minor axis comprised between
b' and b", hence Y2 = b[l - (x/a)2) 1/2. Equation (5.79) becomes
-0
g2 J K2[X'Y'~' Y(~)] d~ = u(x,y) - u 1 (x,y); x,yeOo (5.80)

o
where the second term on the right-hand side is

o
UI (x,y) = J K J (x,y,~) g) (~) d~

-0
5.3.4 Synthesis of material parameters

In this Section a synthesis problem of an another kind will be formulated. Let
us consider once again the inverse formulation of the Laplace's boundary
problem with Dirichlet boundary conditions, but let the inversion be different
from that discussed in Section 5.3.1. Thus, assume an electrostatic field in
the 3D region Q with boundary r, as shown in Fig. 5.9. Suppose the electric
charge density is absent in Q and the unique field source is the given boun-
dary potential <Pf = <p(P)Pef' We wish to obtain the prescribed potential
<Po =<p(P)PeO inside the region Q by forming a suitable distribution of the
permittivity s = s(P)PeO'
Starting from eqn (1.41) for the field that is deprived of electric charges we
have the following equation
diu (s glad cP) =0 (5.81)
Substituting the common distribution of the given (jif and the prescribed <Po for
 • grad w = V2 (x, y, z). Equation (5.82) is linear
with the functional coefficients -gradxq> =EAx,Y, z), -grady q> = Ey(x,y, z),
-gradzq> = Ez(x, y, z), occurring at the differential terms of w(x, y, z), Le. at
ow/ax, ow/oy, ow/oz respectively. The right-hand function is the Laplacian
of q>.
Therefore, the problem of synthesis of environmental parameters leads
to a differential formulation. A practical use of this equation may concern
the one dimensional variability of q> rather than the case of the whole 3D
variability of it that seems to lead to difficult numerical algorithms.
204 Field Synthesis
5.4 SOLUTION OF SYNTHESIS PROBLEMS

5.4.1 General remarks
Many design issues pertaining to the electrical technology are, in fact, pro-
blems of synthesis. The way in which satisfactory solutions are obtained can
be classified as:
• methods based on the theory of optimization,

• methods related to the inverse formulation of appropriate differential
equations describing the electromagnetic phenomena.
Optimization techniques will not be discussed here (but see Section 7.7
later in this book). A fundamental approach to the solution of synthesis
problems formulated with the use of integral equations, can be associated with
the solution methods of Fredholm's and Volterra's equations. Some numerical
problems concerning these methods are presented in this Section. The majority
of the inverse problems of electromagnetism are ill-posed, see Section 5.4.4,
which leads to unstable numerical solutions. Many problems have no unique
solutions. Because of this, particular procedures must be applied in order to
obtain the numerical solutions of such problems. The aim of these procedures
is to stabilize the results of solutions and, moreover, to attribute a physical
relevance to the numerical solutions, which otherwise could be formulated in
a different way. One of the most promising approaches is the technique oj
regularization which is the subject of Sections 5.4.6 and 5.4.7. To apply this
technique, suitable computers are necessary. The simplest method is to use
various methods of approximate solution of algebraic equations. Some of
them are presented in Sections 5.4.2 and 5.4.3.
5.4.2 Quadrature
Let us consider eqn (5.67) for a 20 region, presented in the form
JK(P, s) g(s) ds = u(P), Peno (5.83)

r
It is the synthesis equation for the potential U(P) in no created by the
boundary potential g(s) on r. Suppose g is continuous and limited. It follows
that eqn (5.83) can be discretized. This means that this equation can be
replaced by an operation over the finite set of the values of g(s) related to the
finite set of the chosen U(P) values. This discretization is a quadrature.
We designate, for simplicity, the expression under the integral sign of eqn
(5.83) as h(s) and divide the whole r into m parts (arcs, straight-line sections
etc.). The division points are sv' where v = 1,2, ..., n. Thus, the integral in
eqn (5.83) can be approximated by the sum
n
Jh(s)ds= ~ avh(sJ +R (5.84)
r v=1
Solution of Synthesis Problems 205
z,(
E(P)=?
x, ~
y,7]
Fig. 5.9 Forming the distribution of the permittivity
where h(sy) denotes the values of h(s) at Sy, a y are the numerical coefficients
termed weights of a quadrature, and R is the residue expressing an inaccuracy
due to replacing the integral by a sum; when a 'dense' discretization is made
then R ... 0 is expected.
The synthesis controlled subdomain no can also be discretized by the
arbitrarily chosen points P ~, where I.l = 1, 2, ... , m. In consequence, the
integral equation (5.83) can be substituted by the following system of algebraic
equations
n
2: A~ygy = u~,
y=1
I.l = 1,2, ... ,m (5.85)
where A~y = a y K(P~, Sy) are coefficients of a set of equations being the
products of the weights of quadrature and the values of kernels, gy = g(Sy)
are unknown values of the boundary function g(s) on r, and u~ = u(P~) are
given values of u(P) at the points P~ in no. A set of equations (5.85) can be
expressed in a matrix form
(5.85a)
When m = n is chosen, then [AllY] becomes an n X n square matrix, but
this does not automatically mean that the solution of eqn (5.85a) will be
reached. In practice, if r is a boundary or source line, two types of quadrature
can be effectively applied: the trapezoidal and Simpson IS rules.
For the curved line r l limited by two points Sa and Sb' where s is a curved
line or coordinate on r I' the quadrature can be done by introducing the
division of r I into n - 1 curved elements of the same length liS. When the
trapezoidal rule is applied n may be any real number but using the Simpson's
rule an odd number is recommended. When the division points Sy are
206 Field Synthesis
denoted together with the limits point as: SI = sa and Sn = Sb then the weights
of trapezoidal rule are given by
= liS {
I, if v * 1 or v * n
Oy 1
2' if v = 1 or v = n (5.86)
and for Simpson's rule by
if v = 2, 4, , n - 1
{ 1,
~
3'
Oy = tJ.s if v = 3,5, , n - 2 (5.87)
3, if v = 1 or v = n
When f is a closed line then any point So can qualify as the first point as
well as the last So =SI =Sn' In this case we have for the trapezoidal rule
(5.88)
and for Simpson's rule
Oy = tJ.s {!'
3,
if v = 2, 4,
if v=I,3,
,n - 1
,n (5.89)
The discretization by quadrature can also be effectively applied to functions
of two variables. Consider eqn (5.65) and suppose the boundary conditions
are to be determined on the boundary surface f l C f. If f l is a rectangle
with corner points (sa' ta)' (Sb, ta), (Sa' tb), (Sb, tb) and if the coordinate
system on f l is rectangular, the following Fredholm's equation holds
! !K(P,s,t)g(s,t)dsdt=ii(P)
Sb Ib
(5.90)
So to
The 2D quadrature can be introduced on the above rectangle fl' The

simplest way is to use a uniform division with the step h for both coordinates.
Denoting the coordinates of division points by
Sk for k=I,2, ... ,M; t,for 1=1,2, ... ,N
where M and N are odd numbers and SI =sa' Sb =SM, t l = ta, tb = tN respec-
tively, we have for the trapezoidal rule
2h 2 { 0, if k, I = odd numbers
Ok' =""3 1, if k(or l) = odd number and I(or k) = 1 or N(or M)
2, for any other k, I (5.91)
and for Simpson's rule
Ok' =
h{4,
2
900 8,
if kl = odd number
if kl = even number and k + I = odd number (5.92)
16, if k, I = even numbers
where
4' for corners

(0 = 2, for sides
{
1, inside
The integral over the function of two variables h(s, t) can be substituted by
the double sum formula
Sb Ib M N
f f h(s, t)dsdt = ~
Sa la k=1 1=1
~ aklh(sk, tl ) +R (5.93)
When the controlled points P II , ~ = 1,2, ... , m are chosen in 0 0 then

eqn (5.93) can be written as the following system of algebraic equations
~=1,2, ... ,m; k=1,2, ... ,M; 1=1,2, ... ,N

(5.94)
where
A lIkl = akIK(PII, Sb tl )
This set of equations can easily be transformed into a matrix equation if the
indices k, I are replaced by one common index v = 1, 2, ... , n = M x N.
2 3 4
19 20 21
Fig. 5.10 The synthesis of scalar potential by searching the boundary conditions
208 Field Synthesis
5.4.3 Direct discretization

In many practical problems the explicit form of the kernel is not easy to find.
In such cases the region under consideration may be discretized directly. This
means that a finite element method may be used, but it has to be modified
to suit the synthesis formulation. Let us give an example of such a procedure
which transforms a matrix equation from a form used commonly in analysis
into a form suitable for the inverse problems.
A region 0 is shown in Fig. 5.10, in which Laplace's equation holds. The
region is discretized by using 28 triangles that form 21 nodes. Let the synthesis
problem be to obtain three prescribed values of potential UI2"'" UI4 at
points 12, ... , 14 by determining five values of boundary potentials gl' ... , gs
at boundary points 1, ...,5. Suppose seven values of boundary potential
gIS' ···,g21 are given at nodes 15, ...,21. It is clear that six values of poten-
tial U6' ... , UII at inner points in 0 may assume any values.
Let us start from the general matrix equation in the form that we would
have had by formulating the analysis problem using finite elements. We will,
however, partition the matrices in accordance with the formulation of the
synthesis problem
U6 gl
[A 1] 6x6 [A2]6X3 [Bl]6XS [B2]6X7
UII gs
-_ ..... -------- ------------ = ---- .. ------- ------------ (5.95)
UI2 gl2
[A2]JX6 [A3]3X3 [B3]3XS [B4]3x7
UI4 g21
The next step is to rewrite the above equation in a decomposed form as a
system of two matrix equations with respect to the potentials U6' ... , UII and
UI2"'" UI4
[AI ]6X6[ ~6] + [A2]6X3[::'] = [Bl ]6XS[~I] + [B2]6X7[~:2]

U II UI4 gs g21
[A2]JX6[~6] + [A3]3X3[:,12]
U
= [B3]3XS[~I] + [B4]3X7[~:2]
gs g21
(5.95a)
UII I4
The potentials U6"'" U II are not the subject of synthesis in our problem
and they can be eliminated from the matrix equation. After multiplying the
first of the above equations by [A 1] 6"~ 6, it is then solved with respect to
U6' ... , UII' These potentials are now substituted into the second matrix equa-
tion and we obtain the following equation governing the synthesis problem
Solution oj Synthesis Problems 209
(5.96)
where
[B5]3XS = [A2]jX6' [Al]6~6' [Bl]6XS - [B3]3XS'
[B6]3X7 = [B4]3X7 - [A2]jX6' [Al]6~6' [B2]6X7'
[A4]3X3 = [A2]jX6' [Al]6~6' [A2]6X3 - [A3]3X3'
Thus in our example the problem of finding the unknown boundary
potentials gl,"" gs is reduced to a system of three equations with five
unknowns, as expressed by the matrix [B5hxs with three rows and five
columns. An infinite number of solutions is possible and each is represented
by a set of five potentials gl, ••. , gs with a linear dependence between them.
There are no mathematical reasons to choose any particular solution and the
choice has to be made by considering the physical properties of the system
under consideration.
Suppose gl, ... , gs are the chosen potentials on the boundary. Then the
whole set of 12 boundary potentials gl, ... ,gS, gIS, ... ,gZI forms an
acceptable solution for all inner potentials. The potentials at points 12, ... , 14
will have the values U12,"" U14, that are the best approximations for the
required values U12 , ••• , UI 4 from the engineering point of view. The values of
the potentials U6 ••• U II are not crucial, but may nevertheless be determined in
Fig. 5.11 The synthesis of vector potential by inner current densities

210 Field Synthesis
order to have a complete picture of potential distribution for the structure

given in Fig. 5.10.
Application of such algorithms faces two practical difficulties. The first is
due to the necessity of matrix inversion (e.g. matrix [A 11 6x6 )' In problems in
which no is significantly smaller then n, such matrices tend to be large and
lead to difficulties of a numerical nature. Secondly, a special node numbering
technique must be employed which is usually different from the numbering
provided by most finite element codes used for analysis, see Fig. 5.10.
A similar approach may also be adopted to the synthesis of internal field
sources in a 2D region n shown in Fig. 5.11. In the region under consideration
Poisson's equation is valid for a z-component A z of the magnetic vector
potential A. A system of potentials A\2' ...,A14 is to be generated by deter-
mining appropriate current densities SI"'" S12' Suppose all boundary
current densities at points 1, ... ,5 and 12, ... ,21 are zero and, in addition
let, 513 , ••• ,528 be known. The basic matrix equation is
A6 SI
[A 1]6x6 [A2]6X3 [Cl]6xI2 [C2]6X16
All SI2
------------- ------------ = -------------- --------------
A\2 513
[A2]jX6 [A3]3X3 [C3] 3x12 [C4] 3xI6
AI4 528
It is clear that a similar operation to that applied in the previous example will
lead to a matrix equation of a form analogous to eqn (5.95a), if we replace
[Bl]6XS by [Cl]6x12' [B2]6X7 by [C2]6X16' [B3hxs by [C3hxI2 and [B4hx7 by
[C4hx16' and
The synthesis equation will have the form of eqn (5.96), but there are now
12 unknown variables and only three equations. The problem is thus strongly
under-specified and ill-posed and may be solved using various special tech-
niques discussed in Sections 5.4.4 to 5.4.6.
5.4.4 Conditions for solutions

The synthesis problem can be expressed in a generalized form as
Ag =u (5.97)
where A may be either an integral midcontinuous operator
A =JK(P, Q) ... dQ
or a matrix
The operator transforms the space of solutions g into the space of given func-
tions 'U
A: g -+ 'U
where geg and ue'U.
The problem formulated by eqn (5.97) is correctly posed in the Hadamard
sense [5.15J if three conditions are satisfied:
(1) for each U e 'U the solution g e g exists,

(2) the existing solution is unique,
(3) the problem is stable on the pair of spaces ('U, g), i.e. for any £ > 0
there is such 0(£) that for any two solutions gl' g2 e g and uI , U2 e 'U
we have
U'U(U.,U2) ~ £ ~ U9(gl,g2) ~ O(E)

In design synthesis the second condition may not be necessary; in identifi-
cation synthesis, on the other hand, the uniqueness is strictly required. The
majority of synthesis problems do not satisfy the third condition of the
solution stability. As a consequence, in numerical treatment of the problems,
even very small disturbances in the value of the ii vector, or small inaccuracies
in numerical calculations of the matrix A, may lead to large changes in the
solution vector g.
Most synthesis problems are in fact ill-posed in the Hadamard sense. The
reasons are of two-fold origin: they are either due to the physical nature of
the problem or are caused by inaccuracies introduced by the numerical treat-
ment of the solution. We shall classify such cases for a discretized equation
of the form
n
~ Allvgv = ii ll J.1 = 1,2, ... ,m (5.98)

v=l
(1) m > n - there is more information than unknowns; the problem is

over-specified,
(2) m = n, det[AllvJ = 0 or det[AllvJ "" 0 - the problem is singular or
quasi-singular, respectively,
(3) m < n - not enough information for the given number of unknowns;
the problem is under-specified,
(4) values of ii are approximate, e.g. they are taken from a measurement,
(5) calculated coefficients A llv are not accurate, e.g. due to discretization
error.
212 Field Synthesis
Many physical and mathematical problems are ill-posed in the Hadamard

sense. The well known example is Laplace's equation with Cauchy's boundary
condition.
When a quasi-singular matrix [Allv)nxn is of rank n, its solution may be
found. The probability of reaching the satisfactory solution can be estimated
using a quantity that we will term the index of conditioning, defined as
(5.99)
The smaller the value of cond(A), the better the conditioning of the problem
under consideration. The index of conditioning defined by eqn (5.99) can be
directly applied only when the operator A is inversible. Hence [A l v ) must
be square and non-singular. Definition (5.99) describes a general way of
addressing the problem of quality of conditioning irrespective of the norm
applied, although the numerical value of cond(A) depends on the choice of
the norm.
The following norms are used in order to evaluate the ill-conditioned
systems of equations reduced to a square n x n matrix of the rank n, i.e.
A = [A l v ), where ~,v = 1,2, ... , nand detA O. '*
The first norm:
n
I max L: IAllvl
IIAII = Io;;;vo;;;n (5.100)
11=1
is equal to the maximum sum of the absolute values of elements in one column
of the matrix.
The norm of maximum:
n
max L: IAllvl
IIAII", = IO;;;lIo;;;n (5.101)
v =1
is equal to the maximum sum of the absolute values of elements in one row
of the matrix.
The Euclidean norm:
n n
L: v=l
IIAIIE = 11=1 L: IAllv l2 (5.102)
is equal to the square root of sum of squares of the absolute values of all
matrix elements.
The second norm also called the spectral norm:
IIAlb = max (.JA~)
l~"u::;n
(5.103)
is equal to the square root of the greatest eigenvalue A~ of the matrix

I! = A H • A, where AH is the conjugate (in the Hermitian sense) to the matrix
A, Le. for a real matrix A, B = AT . A.
If a square matrix A can be diagonalized and, in particular, if it is

symmetric, then
(5.104)
where 1.." are the eigenvalues of A. Therefore, the norm IIAI12 is, in this case,
the greatest eigenvalue of A. In consequence, the eigenvalues of diagonalized
matrices A -I are the reciprocals of the eigenvalues of A and the definition
for the index of conditioning (5.99) can be simplified to
cond 2 (A) = An (5.105)

Al
where An and 1..( are the maximum and minimum eigenvalues of A,
respectively.
Example
Let us consider two matrices
[AI] = [~ ~:~} [A2] = [1 0,5 ]

2 1,01
where det[A 1] = 0.1 and det[A2] = 0.01. The inverse matrices are
50
[AI] -I 11
= I -20 -51
10; [A2] -1
= I 101
-200 -100 1
Although the matrices [A 1] and [A 2] are not symmetrical, nevertheless

they can be diagonalized because both of them have two real eigenvalues:
1..( = 0.04875 and ~ = 2.0512 for [AI] and 1..( = 0.004987 and 1..2 = 2.0050
for [A 2] . It is easily shown that the matrices [A 1] -1 and [A 2] -I have the
eigenvalues equal to the reciprocals of the above given eigenvalues and
the matrices [Bl] = [Al]T[Al] and [B2] = [A2]T[A2] have the eigenvalues
equal to the squares of these eigenvalues. The indices of conditioning deter-
mined using the second norm can be calculated following the definition
(5.105)
2.0512
cond 2 [Al] = 0.04875 = 42.08
and
2.0050
cond2 [A2] = 0.004987 = 402.05
The indices of conditioning calculated using the first norm, see eqn (5.100),
are:
cond I [A 1] = (1 + 2)( 11 + 20) = 93
cond] [A2] = (1 + 2)( 101 + 2(0) = 903
214 Field Synthesis
Calculation of condoo[AI] and cond oo [A2] is left to the reader. We notice,

finally, that the index of conditioning for the Euclidean norm (S.102) is
conddAl] = -J( 12 + 0.S2 + 22 + 1.1 2) • -J( 11 2 + S2 + 202 + 102) = 64.6

and
conddA2] = -J (1 2 + 0.S2 + 22 + 1.01 2) . -J( 101 2+ S02 + 2002 + 1002) = 627
However, the use of eqn. (S.99) for examining a priori the solvability of the
system of ill-posed equations is in practice limited. Application of the formula
requires that matrix [A l1v ] is inverted but, whether it can be inverted is
precisely the question we seek to answer.
For the rectangular matrices that are singular or quasi-singular, the basic
equation
1J.=1,2, ... ,m; v=I,2, ... ,n (S.I06)
can be transformed by pre-multiplying by [A l1v ]T into
w=I,2, ... ,n (S.107)
where [Bow] = [Awv]T . [A l1v ] is a square n X n matrix and the definition of
eqn (S.99) can then be directly applied as
cond ([B wv ]) = II [Bwv ] + 11·11 [Bwv ] II (S.108)
where [Bwv ]+ is the inverse matrix of [Bwv ], i.e. [Bwv ]+ = [Bwv]-I. The matrix
[Bwv ]+ is termed the pseudo-inverse matrix to the primary matrix [A l1v )'
5.4.5 Least-squares methods
Basic approach
Systems of algebraic equations where m > n, as for example eqn (S.8S), have
been solved in the past using Gauss' method known as the least-squares
method (LSM). This method provides the foundation for new approaches
developed recently and generally called the regularization techniques.
If the classical Gaussian form of the method is applied to an ill-posed
generalized eqn (S.97), it must be supposed that the solution space 9 and the
space ofgiven function 'U have metrics, and therefore a 'distance' between two
functions is determined. For example the distance between two solutions gl'
and g2' say OO(gl - g2), may be expressed by a real number called the norm
of the function g in the space 9 and it is designated Ilgl - g21b Le.
Odg!,g2) = Ilg! - g2110 (S.109)
If g is the exact solution of eqn (S.97) then the norm vanishes
\IAg - Ull'll = 0 (S.IIO)
When an exact solution g of eqn (S.97) cannot be obtained, e.g. because the
equation is ill-posed in the Hadamard sense, we instead seek an approximate
solution g such that the square of the norm is minimum
\IAg - ull~ = min

geg
(5.111)
as a consequence the square of the norm can be recognized as a functional
mt[u.g] = !IAg - ull~. (5.112)
and a suitable procedure for finding the solution g could be applied by

minimizing this functional in the space 'U.
When eqn (5.97) is discretized to the form (5.98). the spaces 'U and g are
recognized as the Euclidean spaces: g == CRn (n-dimensional Euclidean space)
and 'U == CR m (m-dimensional Euclidean space). In this case the functional
(5.112) becomes
(5.113)
The Euler-Lagrange equation minimizing the functional (5.112) can be

directly obtained from eqn (5.97) by multiplying both sides by the operator
A·: 'U -. g. conjugate with respect to A.
A* Ag = A*ii (5.114)
In the case of discretized eqns (5.98) the Euler-Lagrange system of equations

has n unknown variables
m n m
~ ~ A"'IIAllygy = ~ A"'lIu lI , ro=I.2 •...• n (5.115)
11=1 y=1 11=1
Using a matrix notation we have

(5.116)
The system of eqns (5.115) is called the normal system of Gauss' equations.
If a function g is the solution of Euler-Lagrange's equation (5.114) then this
function is seen as an approximate solution of the i11-posed eqn (5.97) obtained
by the least-squares method. Similarly. if the system of eqns (5.116) is solvable
or. what is equivalent. the matrix
[B]nxn = [A]~xm' [A]mxn
is inversible, then the set of values gy for v = 1. 2•...• n is the approximate
solution of the i11-posed system of eqns (5.98).
The matrix [Al nxn is inversible when the rank of the primary matrix
[Al mxn is equal to the lower of two numbers m. n i.e. when the following
relation holds
rank ( [A ] mX n) = min (m. n) (5.117)
It holds in the case when no linear dependence exists between any two rows
or any two columns of the matrix.
216 Field Synthesis
Example
Consider an over-specified system of three equations with two unknown
variables written in a matrix form
The reader is encouraged to make some guesses as to the possible solutions

by ignoring one equation at a time. If the third line of the matrix is ignored
the solution reads: x = 1, Y = 2; without the second line we have x = 2, Y = I;
and ignoring the first line gives: x = 5, Y = O. If we use the least-squares
method we obtain
[ 1IJ r
1
123
i
12
I 3
l = [36
Le. the normal Gauss's equation has the form
which gives the solution x = 21. y = 1. We can verify the square of the
Euclidean norm (5.102) applied to the above normal Gauss' equation, Le. we
express the sum of the squares of differences between the left-hand and
right-hand sides of this equation.
(2t +I - 3)2 + (21 +2- 5)2 + (21 +3- 5)2 =i
The method guarantees that the value 2/3 is the smallest of all values given
by the combination of pairs x,y satisfying the basic over-specified equation.
To verify this we calculate the squares of the norm for each solution
for x = 1 and y = 2: (1 + 2 - 3)2 + (I + 4 - 5)2 + (1 + 6 - 5)2 = 4 > 2/3

for x = 2 and y = 1: (2 + 1 - 3) 2 + (2 + 2 - 5) 2 + (5 + 0 - 5) 2 = I > 2/3
for x = 5 and y = 0: (5 + 0 - 3)2 + (5 + 0 - 5)2 + (5 + 0 - 5)2 = 4 > 2/3
In many design synthesis such solutions are sought with additional condi-
tions imposed. They may concern, for instance, the desired structural sym-
metry of the object under consideration. The solution by the least-squares
method, while minimizing the Euclidean norm, does not of course promise
to fulfil any such extra conditions. In order to include these conditions, a
suitable enlarged system of equations has to be solved. The following example
shows a suitable procedure.
Example
An over-specified system of four algebraic equations with three unknown
quantities is given
11 21 31] [x] 6.2

[3.1]
1 3 4 Y = 8.3
[
1 4 6 z 11.2
and an additional condition y = z is imposed. First, we enquire into the

solution without the extra constraint. The normal Gauss's equation is
4 10 14]
10 30 43
[x]y = [28.8]
85.2
[
14 43 62 z 122.1
with the solution x = 1.55; Y = 1.0; z = 0.9. The square of the Euclidean norm
is equal to 0.01. We shall introduce the condition y = z and we notice that this
condition does not strongly contradict the approximate solution for y and z.
After adding the third and the second columns of the matrix we obtain a new
system of four equations with two unknown quantities
12] [
~[ ~ [;] = ::;
3.1]
1 10 11.2
the normal Gauss's equation of which is
4 24J
[ 24 178 Y
[x] - [28.8J
207.3
and leads to the solution x = 1.1118; Y = 1.0147. The square of the Euclidean
norm is 0.01265 which is not considerably higher than the 0.01 determined for
the system without the additional constraint.
Singular value decomposition

Recently, the least-squares method has been enhanced and improved by the
use of a special decomposition of the rectangular m x n matrix [A~v], see
Sikora et af. [5.61]. We consider once again, see Section 5.4.4, the square n x n
matrix [B",v] obtained from [A~v]:
(5.118)
The single value decomposition consists of replacing the matrix [A~v] by the
product of three matrices
218 Field Synthesis
(5.119)
The matrix [UIIII ,) is an orthogonal m x m matrix, the columns of which
are the eigenvectors of the transposed matrix [Byw)T; the matrix (Vw') is
an orthogonal n x n matrix, the columns of which are the orthonormal eigen-
vectors of [B wy ); the matrix [SII'y') is the diagonal matrix containing not
more than n (if n < m), and in practice k < n, non-zero diagonal coefficients,
or containing not more than m (if n > m), and in practice 1< m non-zero
diagonal coefficients:
°2
(5.120)
[0]
where
or
The coefficients OJ of [SII'y') are called the singular values of [AlIvl. They
can be obtained as the square roots of the eigenvalues of the matrix [Bwv )
given by eqn (5.118).
The solution of the matrix equation by the use of the least-squares method
with the singular value decomposition leads to the following result
(5.121)
where
and
°2
o
o
[0]
or
[0]
o
o
In the above formulation the given elements Ok (or od denote the smallest
arbitrarily chosen non-zero singular values. If [Sv'~Y has large dimension,
the practical choice of 0k(or 0/) is difficult and thus it becomes the crucial
stage in the procedure.
5.4.6 Regularization
Many authors have tried to reduce the sensitivity of the least-squares method
to the ill-conditioning of the normal Gauss's equation. The principal idea was
to add some stabilizing term 0(g) to the functional mt(u,g), see eqn(5.112).
The first attempt was due to Twomey [5.69], who numerically solved the
Fredholm's equation of the first order by minimizing the sum of the square
of the Euclidean norm together with the second derivative of the solution. The
idea was followed by Philips [5.42] and Tihonov [5.68] who developed the fun-
damentals of this approach, which nowadays is called the method of
regularization.
Following the method of regularization, a special regularized solution is
sought, instead of the solution that cannot be reached directly by the least-
square method. Let us form the functional mrQ[u, g] that consists of two
terms. The first term is the square of the Euclidean norm as expressed by
eqn (5.112); it tends to zero when a good solution is found. The second term
is the square of the norm of a special function, which is related to derivatives
of g. Let the second term be Lg; it has the form
d d
L = Wo + WI dQ + W2 dQ2 + ... (5.122)
where Q denotes the suitable coordinate, for instance Q = ~. The function Lg

belongs to the space OW of the Sobolev's smooth functions.
Hence the functional mrQ[u, g] has the following form
(5.123)
220 Field Synthesis
where « is a chosen real and small number called the parameter of regulariza-
tion. The second term in the functional (5.123) is the functional of
regularization
(5.124)
that can be written as
M
~[g] = «.L: J wi [g(i)]2dQ (5.125)
/= I r
where: Wi are the weight coefficients and g(i) denotes the ith derivative with
respect to the chosen coordinate Q of the solution g(Q). The number M deter-
mines the order of regularization. Numerous synthesis problems have been
solved by restricting the formulation to the first order, Le. with the assumption
= = =
that WI 1 and Wo 0 or Wo 1. The O-order regularization has also been
taken under consideration.
For the Fredholm's equation of the first kind (for instance for eqn (5.63»
the functional mr a can be written as
mra[u(p),g(Q)] =J [JK(P,Q)g(Q)dQ -U(P)]2 dP +

no r
+«J(Wo[g(Q)]2+W t [g'(Q)]2ldQ (5.126)

r
If the discretized form of the Fredholm's equation is given (for instance
eqn (5.85» then mra becomes
a
mr [u l1 ,gv] =11~1 [vtIAl1vgv-UI1]2 + «vtl (WOg~+Wlg~2) (5.127)
where g~ are the values of the derivative of g(Q).

The solution of eqn (5.97) by the use of the regularization method is found
by minimizing the functional (5.123) upon a set of functions g e W. The
Euler-Lagrange's equation associated with this functional is
(5.128)
where
and A· is the integral operator that is conjugate to A.

Let us discuss the problem described by eqn (5.63). If the operator A is
expressed in the form
A = JK(P, Q) ... dQ for PeQo

r
then A * must be
A* = JK(Q,P) ... dP for Qer

00
The combined operator A *. A occurring in Euler-Lagrange's functional

(5.128) has the following kernel
K(Q,R) = J K(P,Q)K(P,R)dP for R, Qer (5.129)

00
and, as a consequence, the Euler-Lagrange's equation, which represents the

synthesis problem given by eqn (5.63) and formulated with the use of the
regularization approach, becomes
JK(Q,R) g(R) dR + a[ wog(Q) - wIg" (Q)] = D(Q) (5.130)

r
where
D(Q) = A*u(P) = JK(Q,P) u(p) dP for Qer

00
Equation (5.130) is an integral-differential equation. To solve it the boun-

dary conditions for g"(Q) should be chosen. They could be formulated in an
arbitrary manner, but suitable in the physical sense for the problem under con-
sideration. In this way the solution can be influenced by designers themselves,
providing they can predict suitable types of solutions of the problem.
The extremum of the functional mr a depends on the chosen value of the
parameter of regularization a. It is clear that for a = 0 we deal with the pro-
blem in its natural, i.e. not regularized, form which leads in many cases, to
an unstable solution. The larger a is the more stable solutions may be
achieved. Moreover, the solutions are also smooth like the corresponding
distributions of the field quantities occurring in physical objects. For very
large values of a the solutions are very stable, but unfortunately larger errors
will be introduced.
If we prefer to use the discretized form of the synthesis equation, for
instance eqn (5.85), Euler-Lagrange's equation for the functional (5.127) is
m n m
2: 2: AWI.LAl.Lvgv + a( wogw -
I.L= I v= I
wIg;) = 2: Awl.Lul.L
I.L = I
(5.131)
where ro = I, 2, ... , n and the second derivative g; occurs in this equation

as the central finite difference of the second order
" gw + I - 2gw + gw - I
gw = h2 (5.132)
where h denotes the step of discretization.

222 Field Synthesis
The matrix form of eqn(5.13I) reads

[A ool1 ]T' [A l1v ] ' [gv] + a(wo[goo] - WI [g:;]) = [A ool1 ]T[u l1 ] (5.13la)
where [g:;] is a band matrix, with the band width consisting of three terms
where [goo] is a diagonal matrix.

All discussion in this section refers to problems formulated by the
Fredholm's equation of the first kind that describes the inverse problem to
Laplace's eqn (5.74). For problems inverse to Poisson's equation, expressed
for instance by Fredholm's eqn (5.74), for determining the unknown source
function f(P) in Q at the given homogeneous Dirichlet's conditions, we have
the following equations
the functional:
mt ll [u(P),f(Q)] = J [J K(P, Q)f(Q) dQ - U(P)] 2 dP +

00 0
+ a JI wo[f(Q)]2 + WI [f' (Q)]21 dQ (5.133)

o
and Euler-Lagrange's equation
J.k(Q,R)f(R)dR + a[wof(Q) - wd"(Q)] = v(Q) (5.134)

o
where
ii(Q) = J K(Q,P) u(p) dP for QeQ

00
5.4.7 Parameter of regularization

Solutions obtained from eqns (5.130) or (5.134) are not unique because they
depend on the parameter of regularization a. The choice of a is a particular
procedure that is associated with the techniques of synthesis. In many pro-
blems, the examples of which we will present in Section 5.5, the value of a
has been chosen by numerical experiment. The idea is to establish an error
function E =f( a) which has a clear minimum. The value of a that corresponds
to that minimum is considered as optimum, and the solution of the regularized
problem at a = aOp! is accepted as the best one.
The error function s, also known as the index of quality, can be defined
as the norm
(5.135)
where gil denotes the approximate solution obtained from the Euler-
Lagrange's equation minimizing the functional mtll[u, g]. When the problem
is discretized the Euclidean norm can be effectively accepted as the error

function
where [g~] is the solution matrix of eqn(5.131a) obtained for arbitrarily

chosen values of n.
Other possibilities of defining the error function are to use either a gradient
of the scalar potential or a curl of the vector potential. When, for instance,
-gradxu = Vx is determined in Qo, then the error function may take the
form
E = Q'U (gradxA~, Vx) (5.137)
A generalized approach to the choice of napt , which does not require a
significant number of solutions to determine the curve E(n), has also been
proposed, see Rudnicki [5.50]. The procedure is called the generalized
discrepancy method. In this method a generalized discrepancy function Q(n)
is introduced into the equation
Ag =u (5.138)
In order to determine E(n), two positive numbers are chosen. The first
expresses the maximum error of the discretized operator, i.e. the discrepancy
of the operator
(5.139)
and, in a similar manner, the second number describes the maximum
discrepancy of the vector on the right-hand side of eqn (5.138):
(5.140)
The value of h depends on the discretization that has been applied. If,
for instance, the quadrature is used, see Section 5.4.2, the residue R occurring
in eqn (5.84) can be accepted as h. The value of 0 has to be evaluated. For
identification synthesis problems the value of 0 is dependent on the accuracy
of the measuring instruments which are used to determine the values of u. For
the design synthesis problems we determine 0 as a value dependent on the
truncation errors of the computer. The generalized discrepancy function can
be described as
(5.141)
It has been shown that Q(n) is monotonous and it has one real root only,
which can be interpreted as n apt ' The solution of a synthesis problem with
the use of the generalized discrepancy method consists of determining this root
in any approximate manner, e.g. using the bisection method. We start by
choosing a rather large value of n = no and then solve the regularized
equation with no as the O-order approximation that leads to the calculated
224 Field Synthesis
value '1(<<0). Next we put «I = «0/2 and compute the solution of the I-order
approximation and calculate '1(<<.). We follow this procedure until '1(<<) takes
a negative value. Finally, we increase « slightly so that '1(<<) = o.
5.5 EXAMPLES OF SOLVED PROBLEMS

5.5.1 Synthesis by searching for boundary conditions
Let us briefly present a problem which was investigated by Sikora and Palka
[5.63J, in order to examine the effectiveness of the regularization techniques
applied to a potential field problem. Let the region 0 be a half-plane y > 0
on which Laplace's equation holds. The boundary r is the x-axis. The synthesis
controlled subdomain 0 0 is a line that is limited by points (X2' y.) = (0.2; 1)
and (X4' Y.) = (0.8; I), see Fig. 5.12. The electric potential cP = V = 1 V is to
be generated by the boundary potential cP(x) =g(x) on r, Le. on the x-axis (the
Dirichlet boundary conditions). The following condition is imposed on the
unknown function g(x)
if x < XI or x > Xs
if XE (xltxs)
at Xl =0 and Xs = 1.
no
Yl ~- - -:-'1- - _ / ' - - - '
I I
· I
I I
· I
I I
· I
I I
· I
I I
I
i: : If
I
· I
~ I
L-. - - ' . _ ' _.~. - ' - r ' - . ' - - - "
xI x2 jX3 x, s
X x'5
Fig. 5.12 Synthesis of the potential on 00

Examples of Solved Problems 225
To use the Fredholm's equation, a surface Green's function for a half-space

(see eqn (5.57» should be taken as the kernel, so that
Xl
JK(X,~) g(~) d~ = V, (5.142)
where:
K(x,~) = 7t [(x - ~)2 + yi] (5.143)
Equation (5.142) can be handled by the I-order regularization at Wo = 0

and WI = 1, Le. the following integral-differential equation is to be solved
Xl
J k(~, ~,) g(~') d~' - ag" (~) = ii(~) (5.144)
with the composed kernel, see formula (5.129)

~ 2
- J Yt dx
K(~, ~,) = 7t2([(x _ ~)(x _ ~,)P + yi[(x - ~)2 + (x - ~')2] + yll
X2
X4
ii(~) = J 7t [(~ _ ~)2 + yi] Vdx

X2
The following boundary condition is chosen for eqn(5.144): g'(xl) =

= O. The error function is defined as
g' (xs)
(5.145)
z.
V mm
0.4
g
t t 0,3
£
0,1
-7 -6 -5 -4 -3 -2 -1 logo{
logot"opt
Fig. 5.13 The error as a function of the regularization parameter

226 Field Synthesis
(V)
g04'()(J
7
6
5
4 ~:::::~~ of:: 2 '10-
6
3 -llIliIIi~--"-~.o(:: 0,6 '10"

2 .
or:: 035'10- 6
6
1 I o(opt :: 0,251' 10-
° 0,3 0,4 0,5 X(m)
Fig. 5.14 Boundary potential - the solution of the synthesis problem
The region under consideration was discretized by using a grid with a mesh
size 0.02 x 0.02 m then eqn (5.144) was numerically solved several times for
different values of the parameter of regularization a. The distribution of the
error function E(a) is shown in Fig. 5.13. The minimum of this function is
2 6
E min=0.17V mm corresponding to a opt =0.251 x 10- •
Figure 5.14 demonstrates the computed boundary potential on the x-axis;
due to symmetry with respect to X3 = 0.5 m only half of the interval (XI, x s)
is shown. The boundary potential function corresponding to aopt is charac-
terized by the largest variation. It yields the synthesized potential shown in
Fig. 5.15 of minimum deviation from the prescribed value 1 V, that reaches
_ (V)
V(x)
Of: 2-10- 6
1,02
./:.--"'- cropt =0.251 . 10 -,
------~-~~-_ ....
0,98
1
.0,96
I
I
I
1"2 )(3
i i i i ~
o 0,1 0,2 0,3 0,4 0,5 X(m)
Fig. 5.15 The synthesized electric potential

a) r,~
R D(S) 1(»
:
I
I
I
I
I
I p(z)
-I 0 I z,S
(b ) r,~
/S(~,S)
R+o
R ----~D(p,5)
I
I
I
I
I
I
I P(z)
-I 0 I z'S
Fig. 5.16 Air-cored coil: (a) with an infinitely thin winding tape, (b) with a winding tape
of finite thickness
3.4070 at the end of the 0 0 line. At a = 0.2 x 10- 6 the solutions become
unstable.
5.5.2 An air-cored coil

Let us present the problem of synthesis of the magnetic field strength in an
air-cored coil. The problem shown in Fig. 5.16 has been investigated by
Adamiak [5.2J and Sikora et al. [5.64J. It has be formulated in two versions,
both postulating the magnetic field strength H to have a unique z-component
Hz(z) = constant on the coil axis along the whole length L = 21 of the coil.
This field is to be achieved by:
(1) the distribution of a d.c. current-sheet J(z) localized on the outer

cylindrical surface of radius R, or
(2) the distribution of a d.c. current density S(z) in the winding tape of
thickness to a.
Both formulations represent an open synthesis problem in which 0 is a 3D
space, 0 0 is the synthesis controlled subdomain confined to a part of the
228 Field Synthesis
z-axis (from -I to l) and Os is the field source subdomain which is a cylinder

of radius " length 21 and wall thickness which is either infinitesimal or equal
to a.
To obtain Fredholm's equation we could use the surface Green's function
for the inside of the cylinder, see eqn (5.61). Another approach is to apply
Biot-Savart's rule expressing the magnetic field strength dHz(z) at the point
P(z) on the z-axis, due to electric current dI = J('e,) d~ of the current ring at
the point Q(~):
(5.146)
Thus, we have for variant (1)
ZE (-I, I) (5.147)
and for variant (2)
L 12~(,2 + '(z _ ~)2P S(~) d,d~ = iI.(z) ,

1 R +a 2
ZE (-1,1) (5.148)
When the first order regularization is applied with Wo = I and WI = I, the

following equations with the complex kernel are obtained:
for case (1)
I K(~, ~,) J(~') d~' + a [J(~) - I K(z, ~)J{.(~) dz

1 1
J" (~)] = (5.149)

-I -I
I K(z,~) K(z, ~,) dz

1
K(~, ~,) = (5.150)

-I
for case (2)

1 1
IKa(~'~')S(~')d~'+a[s(~) -S"(~)] = J Ka(z,~)iIz(~)dz (5.151)

1 -I
Ka(~'~') = J Ka(z,~)K(z,~')dz (5.152)

-I
The boundary conditions that are either J'(-I) = J'(I) = 0 or S'(-I) =

S'(/) = 0 lead to the solutions presented in Figs 5.17 and 5.18. It can be seen
from Fig. 5.17a and 5.17b that the maximum magnetic field uniformity of
about 99% is achieved at a = 10- 9 • Similar results apply to the coil of finite
thickness tape shown in Fig. 5.18a and 5.18b. In both cases the necessary
distribution of current must be precisely followed. In practice, the coil has to
(a) 15 r---,--r----r------.
Ci = 10- 9
10 t---+---t---+-~
-
E 5 t---t-::I~'--+---+--~
<{
- 5 t-----+---+---\-~I-.----l
(b) 3
~
2
1
/ r\
/ /..
--
~ ~
1.0 z
0 ~
\~ Of«==10"
~.
~ ./
:.. ·1 10"
"0
·2
V" \ ' d: 10"
-3
\
-4
\
Ci. = to,S
-5
Fig. 5.17 (a) Current-sheet distribution of the coil with an infinitely thin winding tape,
(b) relative error of the magnetic field strength
be partitioned into a number of sections, each of them supplied by a current

of some average value which is determined from the distributions shown in
Figs 5.17 or 5.18.
5.5.3 Boundary conditions for a transmission line

In this Section the synthesis problem inverse to the diffusion equation will be
presented following the paper by Sikora and Palka [5.62]. Let us consider the
transmission line equation for the voltage u:
230 Field Synthesis
(a) 4,0 r - - - - - - , - - - - - - - , C( =10- 9
3.0 t - - - - - - t - - - - - - H
....E
<l::
II)
- 1,0 t----~~-I__-~
- 2.0 L - -.L.. ----l
(b) 3,0
2,0
1,0
1,0 z
0
0
;--
--20
":::I '
-3,0
-4,0
-5,0
-6,0
Fig. 5.18 (a) Current-sheet distribution of the coil with a finite thin winding tape,
(b) relative error of the magnetic field strength
a2u 1 au (5.153)
ax 2 = ~2 at
where x determines the distance from the beginning of the line, t denotes the
time and B = l/.J(RC) where Rand C are the resistivity and capacity of the
line, respectively, both expressed per unit length of the line.
The following identification synthesis problem is considered. Suppose at
the distance XI the time-dependent voltage u(x., t) = u(t) has been measured.
The problem is to determine the Dirichlet boundary conditions u(O, t) = g(t)

at the beginning of the line, with the assumption that the initial conditions
for the whole line are zero, Le. u(x, 0) = O. The complete formulation of the
problem is therefore as follows:
(1) the given voltage function at the point XI is U(XI' t) = u(t) for t ~ 0,
(2) the initial conditions are u(x,O) =0 for x ~ 0,
(3) the boundary conditions: u(O, t) = g(t) for t ~ 0 should be determined.
The problem is described by Volterra's equation of the first kind

t
JK(t, 't) g ( 't) d't = U(t) (5.154)

o
with the kernel
K(t, 't) = 213~[1t~: _ 't)3] exp [413(~~ 't)].

having the singularity at t = 'to Equation (5.154) can be reduced to the
equivalent Fredholm's equation of the first kind
T
J K (t, 't) g ( 't) d't = U(t) (5.155)

o
with the kernel
for 0 ~ 't < t
for t ~ 't'~ T
(V)
got (t} 1) ct :
10. 2
2)(t : 10. 5
4 3)C! = 10. 4
4) 0: : 5'10"
o 0,2 0,4 0,6 0.8 1,0 tIs)
Fig. 5.19 Time-dependent function of the voltage at the beginning of the line
232 Field Synthesis
(V)
(i0/'ftJ
~2
1,0
0.8
0,6
0,4
0,2
o 0.2 0.4 -0,6 0,8 1,0 tis)
Fig. 5.20 Time-dependent function of the voltage at x}t l1(t) is the measured function, 1-4
are recalculated functions
u(t) = {2t/T,
1,
Equation (5.155) has been discretized and then numerically solved for the
following data: R = 10m- I , c= 1 Fm- I , XI = 1m. Figure 5.19 shows the
voltage u(O, t) = ga(t) traced at four chosen values of the regularization
parameter, and Fig. 5.20 presents the recalculated voltage function ua(t) that
is expected to be close to u(t). It can be seen that at a = 5 x 10- 6 the
function ua(t) (curve no. 4) approaches closely the u(t) and, in consequence,
curve 4 on Fig. 5.19 is the best solution for the voltage function at the
beginning of the line. For smaller values of a the function ga(t) becomes
unstable and this is the main feature of an inverse problem that is ill-posed
in the Hadamard sense.
5.5.4 Other problems

For an increased treatment of solved synthesis problems pertaining to
electrical technology the reader may wish to refer to the papers published in
majority in the IEEE Transactions on Magnetics and in other journals.
Examples oj Solved Problems 233
An identification synthesis is reported by Iwamura and Miya [5.18] and

Lord et al. [5.24]. A shape design of an electromagnet is presented by
Armstrong et al. [5.4], Arumugam et al. [5.5], Di Barba et al. [5.8], Marrocco
and Pironneau [5.25], Nakata and Takahashi [5.30] and Siebold et al. [5.58].
A design of various electric devices is reported by Preis et al. [5.43],
Ratnajeevan and Hoole [5.48], Rudnicki et al. [5.50], Russenschuck [5.52] and
Sikora [5.59]. The use of synthesis techniques for a design of NMR
tomography is presented by Gottvald [5.11], Miyata et al. [5.28] and
Mustarelli et al. [5.29]. For an optimization strategy pertaining to the synthesis
problems the reader may wish to refer to the papers by Bellina et al. [5.6],
Grago et al. [5.6]. Gitosusastro et al. [5.10], Gottvald et al. [5.12], Kasper
[5.20], Michalski and Sikora [5.26], Nakata et al. [5.31], Osama [5.32] and
Saldanha et al. [5.54].
6
Coupled Fields
Janusz Turowski
6.1 TYPES OF COUPLED FIELDS. STRONG AND

WEAK COUPLING
By the term coupled fields we understand a system of directly interacting
physical fields: electric, magnetic, mechanical forces and stresses, thermal,
gravitational fields and flow field of currents, liquids and gaseous conducting
media, plasma and similar. Coupled fields also encompass the interaction of
electromagnetic reaction of secondary circuits to primary fields. The latter
is especially evident when changes of forces and torques take place at the
start, reverse, braking, automatic restart and similar transient conditions of
electrical machines. Such transients are usually investigated with the help of
Euler-Lagrange's equation (6.23) using equivalent lumped electromagnetic
parameters - flux linkages A.k' inductances L ki and forces fek' calculated from
field analysis. Typical examples of coupled fields include magnetohydro-
dynamic, magnetogasdynamic and electromechanical levitation fields, radial
traction and axial fields of rotating and linear electrical machines, electro-
mechanical fields of transient electrodynamic short circuit forces, transient
electromagnetic/thermal fields in skin hardening of steel parts, electromagne-
tic pumps etc. Such fields can be considered as having a strong coupling. They
are often associated with the strong eddy-current effect.
In static, quasi-static and slow-varying fields the diferences in time constants
between electromagnetic, thermal and mechanical processes are such that the
coupling is practically in one direction, Le. the thermal or mechanical stress
sources are of electromagnetic origin, but there is little or no feed-back effect
on the primary magnetic field. Such fields, which cover a wide spectrum of
technical problems, are said to have a weak coupling. These fields can often
be investigated independently or regarded as a limiting case for fully coupled
fields, after transients have decayed.
The degree of coupling between physical fields, as well as the distinction
between strong and weak coupling, can be evaluated with the help of Reynolds
number R M (6.74). For instance a definite electromechanical coupling in
the form of levitation of linear motors occurs with high-speed motors at
R M = 10, ... , 25, whereas a noticeable reaction of induced eddy-currents
already exists at R M > 0.5 [6.22].
An electromagnetic field is in its nature the coupling of electric and magnetic
fields, which is observed as the propagation of electromagnetic waves. This
Electro-Magneto-Mechanical Fields 235
occurs at higher frequencies and is beyond the scope of this book. The diffu-
sion process in conducting media, on the other hand, is very much the subject
of our discussion.
The emphasis of this chapter is on a physical and mathematical description
of coupled fields. Once the problem is formulated its solution may be found
using one of the methods discussed in other chapters of this book. Numerical
solutions are usually very difficult owing to the complexity of the problem
formulation. Thus analytical, semi-analytical and empirical solutions are often
sought, after suitable approximations and simplifications to the model have
been made.
6.2 ELECTRO-MAGNETO-MECHANICAL FIELDS

6.2.1 Hamilton's principle
In the analysis and synthesis of coupled fields we concentrate on investigating
the state, motion and energy conversion. In Lagrangian mechanics of charged
particles moving in electromagnetic fields, the state of a dynamic system at
a given instant of time can be described as a point in a 2n-dimensional space
of n generalized coordinates qk = qit) and n generalized momenta Pk = Pk(t)
or corresponding generalized velocities qk = qk(t) [6.34]. This approach is
usually applied to systems with lumped parameters [6.31] or if field effects
can be expressed in terms of discrete parameters 'Pij, Mij, Rij.
The Lagrangian description of interactions of electromagnetic fields with
external field sources of charge density p or current density J, runs in strong
analogy to the description of particles ([6.8], p. 571). The finite number of
coordinates qit) and qit), where k = I, 2, ... , n is substituted here by an
infinite number of degrees of freedom, and the generalized coordinate qk is
substituted by continuous fields <1>; (x). Therefore both Hamilton's principle,
which according to Hammond ([6.7], p. 31) ' ... is more general than the law
of the conservation of energy in its usual form .. .' (see also reference [6.34],
p. 31) and following on from it, as the necessary condition for an extremum,
Euler-Lagrange's equation, refer to electromagnetic fields as well as to discrete
systems with lumped parameters. Both these relationships follow from energy
interactions, whereas the system of generalized coordinates depends on the
method of calculation of these energies (see Section 6.2.4). We shall start our
discussion with the system of lumped parameters.
The conservative (loss-less) kinetic energy T of the system is a quadratic
form of the generalized velocities
n n
T(q, q) =~ ~ P;k(qt> q2' ... , qn; t)q;qk (6.1)
;=1 k=l
In the case of nonlinear media, using the more convenient coordinate sys-
tem (Xk' ik ), it is better [6.34] to employ the so-called kinetic coenergy T',
evaluated for example from magnetization curves for flux linkage ",' = ",'(0
(Fig. 6.1) as
236 Coupled Fields
wm = ,. d ~ "'i ; ., ,
;. I I
~ +---/---''----:::::.-
i,
Io I .
w~ = ' 'I;d i;'
I I
COENERCY
t
.,
/. /.
I I
Fig. 6.1 Magnetic coenergy W~= i;'II; - Wm of the ith element of an electromechanical
system [6.34]
q••.... q. n
T'(q, lI) = f ~ p£(ql' ... , qn; lIt, ... , lI~; t)dlI£ (6.2)
0..... 0 k=1
where primes denote variables of integration.

With the help of the Legendre transformation of coordinates [6.34] we
obtain a simple relationship between the magnetic energy Wm and magnetic
coenergy W~. For the ith element
(6.3)
from which it follows that for linear systems W~ = W m and T' = T. The
potential conservative energy V is linear form of general coordinates q;(t)
n n
V(q) =~ ~ V;k(q., ... , qn; t) (6.4)
;=1 k=1
The most general mathematical description of the transformations and pro-

cesses mentioned above is Hamilton's principle [6.34], which can be derived
for a mechanical system from d' Alambert' s principle and the principle of
virtual work, using the calculus of variations. It comprises not only all kinds
of fields but also electric and magnetic circuits with lumped parameters, the
movement of discrete charges and masses as well as mechanical bodies and
particles.
The Hamilton's variational principle states that the variation OJ of the time
integral of the Lagrange state function, which is also called the conservative
Lagrangian
L(q, lI, t) =T - V (6.5)
must be zero during motion of a system between fixed end states qj(t l ) and
qj(t2 ), i.e.
12
M= 01 L(qj, qj, t) dt = 0 (6.6)

II
This means that motion of the system in time from t l to t2 between the two
end points qj(t l ) and qj(t2 ) is accomplished in such a way that the functions
q(t) forming a functional
12
1= f L(q, q, t)dt (6.7)

II
make this functional stationary (usually a minimum) [6.34]. Under static con-
ditions and only then, this condition reduces to the principle of minimum
potential energy aVlaqj = O.
Since
expresses the average value of energy difference (T - V) in time interval

t2 - t 1 , hence M = (t2 - t l ) o(T - V)mean = 0 and we can state that nature
tends to the equalization of mean kinetic energy with potential energy of a
system during its motion.
In the case of nonlinear dissipative systems (with power losses and energy
sources) we replace the Lagrange function (6.5), also called the conservative
Lagrangian or kinetic potential, with the nonconservative Lagrangian
L FO = (T' + TJ;) - (V + Vo) = L + (TF - Yo>, L = T' - V (6.8)
which also satisfies the condition of Hamilton's principle [6.34], where
IoFdt
I
TF(q) =
is the non-conservative kinetic coenergy of the system (energy of power

losses),
n I
F(q) = ~ - rj(qY
j=1 2
is the Rayleigh dissipation junction, equal to half the Joule's losses (rj-
resistance, qj - current) or half the frictional losses (rj - viscous frictional
resistance, qj - velocity of the body),
Io Gk(t) dq!c
qk
Vo(q) =-
238 Coupled Fields
is the potential energy of non-conservative forces of constraints imposed by

external energy sources, and
is the non-conservative force of constraint applied from the sources, e.g. input
voltages applied at the terminals of an electric network or externally applied
mechanical torques or forces.
From Hamilton's principle many physical laws and principles of mechanics,
electrodynamics, electrothermics, thermodynamics, electrochemistry etc. can
be derived.
6.2.2 Euler's equation

It follows from the previous Section that coupled fields and systems can be
investigated by solving a variational problem of the form
1= ff F(u, Un uy,x,y) dxdy = extremum (6.9)

(l
where the integrand F represents the energy or quantity proportional to it,

contained in the system with n degrees of freedom. Every variable x and y
has a vector in the space Rm [6.15], and u = u(x,y)eRn • Coordinates of the
vector Rn are general coordinates of the system. It is assumed that general
coordinates and general velocities of the system are known at the instant
Xo,Yo and
ueC'[xo,Yo;x,y],
The functional (6.9) has a more general significance than (6.6) and comprises
both Hamilton's principle
and the principle oj least action [6.23]
The two principles have a different physical interpretation [6.23]. In the case
of Hamilton's principle the subject of variation is the path of integration
between two points, which is not changing in space and time, and on this
path the equations of motion and conservation laws are not observed. In the
principle of least action, on the other hand, the end points are fixed in space
- but not in time, and on the path being subject of variation the energy con-
servation law is satisfied.
The task of finding the extremum of the functional (6.9) in region Q can
be resolved with the help of calculus of variation. We are looking for the
surface u(x, y) which satisfies the condition (6.9) and boundary conditions
u(C). Let us assume that
u(x,y) + all (x,y) (6.10)

represents an adjacent, sufficiently near surface on which TI(x, y) is an
arbitrary function satisfying the boundary condition TI(C) = 0 and a is a
small constant, of positive or negative value, which is the parameter of this
surface.
For the surface (6.10) we have
1 + OJ = JJF(u + aTl, Ux + aTlx' uy + aTly,x,y) dxdy (6.11)

n
where
oTl
Tlx = ox'
After expanding (6.11) in powers of the parameter a with the help of Taylor's
formula
2
I(a) = 1(0) + al'(O) + a2! /"(0) + a3 k(a) (6.12)
we see, that the necessary condition of existence of the extremum of the

functional (6.9) is that the first variation OJ must be zero, Le.
OJ = a (aIJ
oa a=O
=0 (6.13)
The analysis of the second variation 52/ = a 2/,,(0), with the help of
Legendre's, Jacobi's and Weierstrass's conditions [6.15], reveals that usually
52/~ 0, which means that the functional attains a minimum.
From eqn (6.13) it follows that
J
II [
oF of
OJ = a n OU 11 + oU Tlx + OUy Tl y dxdy
x
of
=0 (6.14)
The last two components of eqn (6.14) are transformed using Riemann's
formula [6.23]
H[OQax - oPI
n(c) oyj
dxdy = J (Pdx + Qdy)
C(Q)
(6.15)
by substituting
of of
Q = -TI and -P =-11
oUx OUy
240 Coupled Fields
so that
ifa [:x (:~ nJ + :y (::ynJ] dxdy =ifn (n:X :~ + nx :~

d dF dFJ
+ n~ ~ + lly~ dxdy
uy vu vu
y y
and thus from eqn (6.15) we find
In (- dUy
c
dF dx + dF dyJ =ifn (~ dF + ~ dFJ dxdy
dUx n lax dUx dY dUy
+ Jj [:~ nx + :~ nyJ dxdy (6.16)
From the boundary condition n( C) = 0, the line integral in eqn (6.16) vanishes
and after substituting eqn (6.16) into eqn (6.14) we obtain
OJ = aifa n(dF - ~ dF - ~ dFJ dxdy = 0

LdU dX dUx dY dUy
(6.17)
Since n = n(x, y) is an arbitrary function, eqn (6.17) can be satisfied only

if the term within the brackets is identically equal to zero. In this way we
have obtained the necessary condition for the function u(x, y) to mini-
mize the functional (6.9). This condition is known as Euler's differential
equation
dF d (dFJ d (dFJ (6.18)

dU - dX dUx - dy dUy =0
6.2.3 Eurer-Lagrange equation
If we express the integrand in eqn (6.9) in the form of Lagrange's state
function as
F(u, ux' uy,x,y) = L(qk' ilk, t) = T' - V (6.19)
then eqn (6.9) will represent Hamilton's principle (6.7)
12
JL(qk' qk> t) dt = min. (6.20)

I,
= = = =
After the substitution of x t, Y 0, u(x, y) u(x) qk(t) and uy = 0, we
obtain from eqn (6.18) the well-known Euler-Lagrange equation for con-
servative electromechanical systems
dL _.!! (dL J= 0 k = 1,2, ... , n (6.21)

dqk dt dqk '
For an electromagnetic field eqn (6.21) takes the form [6.8]
(6.21a)
where .e = .e(cPi' aa cPi ) is the density of the Lagrangian, whereas 'coordinates'

and 'velocities' are fa and a~ fa respectively; the finite number of k coordi-
nates qkt Qk in eqn (6.21) corresponds to an infinite number of degrees of
freedom and a point x a in the space-time coordinate system. The continuous
fields cPi(X) contain a discrete indicator i = 1,2, ... , n and a continuous
indicator x a; the generalized velocity Qk is replaced by a four-vector a~cPi'
which has the form of a gradient; the action integral (6.7) takes the following
form for fields
(6.7a)
In practical calculations of system dynamics, eqn (6.21) is preferred to eqn

(6.21a) and the emphasis is shifted to finding appropriate equivalent lumped
parameters M ij , Rij (6.62) of the system.
On the other hand, taking into account the nonconservative Lagrangian
(6.8)
(6.22)
we obtain the Euler-Lagrange equation for dissipative electromechanical
systems
-
:~ :t[:~J :~ = -Ok - (6.23)
where k = 1,2,3, ... , n is the number of degrees of freedom of the electrical

and mechanical systems and other symbols have already been explained. Other
equations of electrodynamics can be obtained in a similar way (see references
[6.31; 6.32; 6.33]).
6.2.4 Static electric and magnetic fields in isotropic media

Many technical problems are resolved using static equations
curlH = J, curlE = 0, J = (JE, D = eE, B = ~H, divD = p, divB = 0
(6.24)
In current-free regions (cur/H = 0) we have a potential (irrotational or
conservative) field
E = -grad Ve , H = -grad Vm (6.24a)
The energy of an electric field per unit volume (J m- 3 ) is
Ve = e (aYJ2 (aYJ2 (aYJ2

2 2
ED eE e
we = 2""" = T = D2e = "2egrad2
1
a: + e aye + e az
(6.25)
242 Coupled Fields
In a region V containing a charge with the volume density p, the energy of

the system is
(6.26)
The energy of a magnetic field per unit volume (J m - 3 ) is

2 2 2 2
Wm = HB = J.1H = B =! J.1grad2 Vm = J.1 [CWmJ2 + J.1 [a VmJ + J.1 [a VmJ
2 2 2J.1 2 ax oy OZ
(6.27)
In terms of the magnetic vector potential A defined as B = curl A the energy
of the magnetic field can be expressed as
2
wm=B =~curJ2A=~[[OAz _ OA yJ2 + [OA x _ OA zJ2 + [OA y _ OA x J2]
2J.1 2J.1 2J.1 oy OZ oz ax ax oy
1 1
= 2H curl A = 2 [A· curlH - diu (H x A)] (6.27a)
which follows from the identity diu (H x A) = A . curl H - H . curl A.

In a region containing currents with density J = curl H, the energy of the
power loss per unit volume (W m - 3) is
W·
J
= pP = EJ =-p1 E 2 = oE 2 (6.28)
where resistivity p = 1/0.

The energy in the entire region of volume V is
wm =-21 Li 2 =-21 J H· B d V =-21 J A· J d V (6.29)

v v
Equation (6.29) follows from eqn (6.27a), because
J diu (H x A) dV = J (H x A) dS = 0
V-'oo 5-00
since at infinity H=k(/r 2,A =k2/r,dS=k 3r 2, and thus HAdS=k4 /r-+0.

The potential energy V of a charge Q in an electric field with a scalar
potential Ye, in a region of volume 0, is given by
V H
= Jp Ye dO = QYe
(2
(6.30)
The kinetic energy of a charge Q in a magnetic field described by a magnetic

vector potential A, in a region of volume 0, is
T = JJJ p (A . v) dO = Q (A . v) (6.31)
(2
where v is the velocity of charge Q in the magnetic field of flux density

B = curl A. The energies given by eqns (6.30) and (6.31) are included in the
Lagrangian L = T - V (eqn (6.5», expressed in terms of values of electric and
magnetic fields.
6.2.5 Maxwell's equations as a consequence of Hamilton's principle

In the case of motion of a particle of mass m and charge Q with velocity
v in eqn (6.6), which represents the Hamilton principle, the conservative
Lagrangian (6.5) consists of kinetic mechanical and electrical energy [6.7]
(6.32)
Substituting (6.32) in the Euler-Lagrange equation (6.21), where qk =XI'
q = v, i = 1,2, 3, we have
[1
- a -mv·2 +QA·v-QV - - -
ax! 2 I e
J d [1
a -mv·I2 +QA·v-QVe =0
dt aVi 2
J
d
QV(A'v) - QVYe = dt (mv + QA) (6.33)
From the vector identity ([6.15], p. 90)

V(A'v) = (AV)v + (vV)A + A x (V x v) + v X (V x A)
= (vV)A + v x (V x A)
since the left-hand side of (6.33) is a partial derivative with respect to Xi' and
the operation V on the velocity Vi equals zero. Hence eqn (6.33) takes the
form
d dA 3 aA
Q(vV)A + Qv x (V x A) - QVYe = dt (mv) + Qdi + Qi~ aXi Vi
from which Lawrynowicz [6.15] finds
d
dt P = Q(E + v x B) (6.34)
where:
p = mv mass momentum (6.35)
E = -aA
at- - VV
e
electric fields strength, (6.36)
B=Vx A vector of flux density. (6.37)

The right-hand side of eqn (6.34) represents the Lorentz force, which is thus
given by the time derivative of the particle momentum dp/ dt. After calculating
curl of both sides of eqn (6.36)
curl E = - ata curl A - curl grad ~

244 Coupled Fields
and considering eqn (6.37) and vector identity curl grad Ve = 0, we obtain
from (6.36) the second Maxwell equation
curlE
aB
= -- (6.38)
at
and then, from eqn (6.37) and the vector identity div curl A = 0, we have
divB = ° (6.39)
In a similar way, from Hamilton's principle, the first Maxwell equation and
div D = p can be derived [6.15].
6.2.6 Generalized vector of power density of coupled fields {6.31}

Extending Poynting's theorm using Hamilton's principle and the Lagrangian
state function ([6.34], p.38) to other kinds of energy, one can, by analogy
with energy conservation law, formulate an integral principle applying to a
generalized vector N of power density of the set of physical fields
Ps = , N· dS (6.40)
S( V)
where Ps is the power of physical fields flowing in or out from a space of

volume V and surface S( V), N is the generalized power density vector of
physical fields, where
N = S + U + Qp + qN + qSB + M + Ee + I + Sch + . . . (6.41)
where
(6.42)
is the complex Poynting's vector, which for sinusoidal fields has complex
components
£c = !(E.myH:z - E.mzH:y)
~y = ! (E.mzH:X - E.mxH:z)
~z = ! (E.mxH:y - E.myH:X) (6.43)
and
-divS = ata (ED HBJ + aE2 + pE·v + aE(Eext + v x B)

""2 +""2 p
(6.42a)
u =v dWmech Umov's vector of flux density of mechanical power

dV in elastic media;
qE= -'A.grad9 vector of thermal power flux density dissipated by
conduction (Fourier's law);
vector of thermal power flux density dissipated by
convection (Newton's law);
qSB = Iv Kv (04 - O:mbJ vector of thermal power flux density dissipated by

radiation (Stefan-Boltzman law);
M vector of light radiation power, emitted by a sur-
face (emittance power);
Ee vector of radiant power flux density incident to a
surface (irradiance);
I vector of acoustic power density (sound intensity)
Sch vector of electrochemical power flux density etc.
Vectors qSB' M, Ee can be also expressed by Poynting's vector S, but here
they are isolated to emphasize their role and means of calculation. Equation
(6.40) therefore expresses the first rule of thermodynamics ([6.21], p. 13)
d
dt (Wk + W;) = P ex + Pem
where W k and Wi are kinetic and internal energy, P ex is the power of external
forces and Pem is the flow of electromagnetic energy through the surface S
surrounding the volume V of the body, which according to the energy con-
servation principle means that the increment in time of the kinetic and internal
energy is equal to the power of external forces and flow of electromagnetic
energy.
6.1.7 Field of forces

An elementary force acting on a conductor with current 1= J . dS placed in
a field of flux density B can be calculated using one of three fundamental
methods: Ampere's and Biot-Savart's law, the Lorentz force and the principle
of virtual work. Maxwell's stress tensor (6.66) is also very useful when calcu-
lating local forces from electromagnetic field parameters.
Ampere's and Biot-Savart's laws

Ampere's law is also called the law of flow [6.8] § H . dl = i and the formula
for force F 2 which is the action of the current loop 1 on current loop 2
(Fig. 6.2)
_ J10 J I J2{ 1 dl 2 x (dl l x '12) _ J10 J I J2{ 1 dl l 'd12 (6.44)
F2 - - - 1 1 3 - - - - 1 1 - 3 - '12
4n 12 II r 12 4n II 12 r 12
Using expressions for loop and flux linkage \II we have
_ J10 t I dl l d _ J10 ttl dl l • dl2

A--1 1- an \11--111-- (6.45)
4n II rl2 4n II 12 r12
A different form of eqn (6.44), expressed by the flux linkage \II(x) and current
I, can be obtained by writing
J1 oI1I2
= --grad t t dl l • dl2
F2
11 - - = 12 grad \II
(6.46)
4n II 12 rl 2
246 Coupled Fields
Fig. 6.2 Interaction of current loops
Then, considering eqn (6.45), we have the general expression for components
of force as in the method of virtual work
F = IO",(x) = 1 / oM(x)
x 1 2
OX OX
where
(6.47)
is the mutual inductance of conductors (M12 = M 21 ) without iron. The pre-

sence of iron provides additional sources for A (B = curl A) and the integral
in eqn (6.47) may then no longer be symmetrical. The statement M I2 = M 21
is not in general true when there is iron in the magnetic field.
If conductors cannot be treated as filaments with negligible thickness it is
better to calculate inductance from the field vectors in volume V
L = ;2 HI H . B d V = ;2 HI A . J d V (6.48)
v v
The force acting on element dl of a conductor with current I in a magnetic
field B can be calculated from Biot-Savart's [6.11] law
dF = I(dl x B) = nBld/sin L (dl,B) (6.49)
The force per unit volume (N m -3) is
F 1 =J x B (6.50)
and the force acting on a conductor of length I with current I is
F=/JdIXB (6.51)
1
Lorentz force
Lorentz's force density in N m -3, according to eqn (6.50) is
f L = J x B = J x curl A = p(E +vx B) (6.52)
where p is the volume density of a charge moving with velocity v in field B.

Lorentz's force is therefore
FL = Q(E + v x B) (6.52a)
The total force acting on an arbitrary part of a conductor or winding of
volume V is
F = HI (J x B) dV= HI (J x curiA) dV (6.53)

v v
For sinusoidal time variation and complex amplitudes lm' 11 m , A m the force
can be expressed in the form
(6.54)
with the components

Fxmean = t Re(lmyll:, -lm,ll:y)
Fymean = t Re(lm,ll:x -lmxll:,)
F,mean = t Re (lmxll:y - lmyll:x) (6.55)
Principle of virtual work

The principle of virtual work states that arbitrary displacements dxj of
elements of a winding in the j-direction under the influence of an electro-
magnetic force F ej (Fig. 6.3) are so small, that they cause no changes of
current ik • At the same time the work of the force is equal to an increment
of energy of the system
Fej dxj = d ( W) it = conslanl
The instantaneous force acting on the kth element of a conductor or winding
in direction xj is equal to the partial derivative of energy Wm of the magnetic
field in that direction where ik = constant. From the Legendre transformation
it follows that for nonlinear systems it should be the coenergy W~. Therefore
the force
F. = [OW~J = - [oWmJ (6.56)

eJ OXj h=conslanl OXj 'lit = conslanl
The magnetic energy

B
Wm = JJ H . dB d V
vo
and coenergy (Fig. 6.1)
H
W~ = JJ B· dH d V
v 0
248 Coupled Fields
Fe) ••••
, ..
Gi n
\
I
Fig. 6.3 Virtual work in a system of currents
can be calculated directly from the fields H, B, A and J. Alternatively, self-

or mutual inductances according to eqns (6.1), (6.2) and (6.48) can be
used
Having a field of magnetic vector potential
the integration can be limited only to the conducting regions, where the current
*
density J O.
Examples
(1) Lifting force of an electromagnet (Fig. 6.4)

For simplicity we assume an analytical approximation of the magnetizing
characteristics of the iron core [6.23] in the form
B = cHI / n
For isotropic transformer laminations the constants are typically n = 10,
c=O.9.
The magnetic coenergy W~ in a magnetic circuit is therefore
.
~I
~ t--o+--'"
.5 5
Fig. 6.4 Lifting electromagnet
H H H
W~ = JJ B· H d V = J c J HYen dH d VFe + J J ~OHair dH d Yair

V 0 VF< 0 Vair 0
-
- c 1/n
1
+ 1nn/n+I)/nv
'
1 H2 dT.T -
Fe + 2~o p "air -
0 91 HUV
• C Fe Fe
~OH2air"air
+"2 T.T
~
~O
"2 Hair Yair ="2
2 ~O (INJ
2x
2
4Sx = ~o~S
(IN) 2
According to eqn (6.56) the lifting force is
F- -
[awmJ
ax l=constanf
-
In the where the excitation current is sinusoidal, and the flux cJ> = cJ> sin rot,
the instantaneous force is
cJ>2
F(t) = 4_m (1 - cos2rot)
4~oS
Conclusion
The magnetic pull in Nm- 2 per unit surface of the magnet (cf. Fig. 6.7b)
is
B2 HB
p=-=- (6.58)
2~o 2
(2) Magnetic pole pull

The magnetic pole pull (in N) of one pole of an electrical machine within one
pole pitch 't = 1tDl2p and armature length Ii is
B2
FN = liJ (X} dx (6.59)
2~o
250 Coupled Fields
(3) Magnetic attractive force

The magnetic attractive force of field B, parallel to the steel surface (Bn = 0)
is normal (cf. Fig. 6.7c) to the surface ([6.30], p.368), but much weaker
p~--
B7 (6.60)
21!,l!o
6.2.8 Vibration of busbars and conductors

Instantaneous electrodynamic forces, together with local excessive heating
brought about by strong leakage fields, are a significant factor in decreasing
the reliability of large electrical machines, power transformers and other
electric power equipment in power systems, power stations, substations,
electric furnaces etc. Particularly important is the calculation of these forces
and tensions during transient short-circuit conditions and other transient
faults, such as self-operated switching on, impulse overloading, etc.
Nowadays there is plentiful literature on the calculation of electrodynamic
forces in transformers in quasi-stationary states which neglect changes of
self- and mutual inductances Mij during winding vibration. Less researched is
the influence of changes of Mij in the course of vibration of busbars and
windings. This effect was widely investigated [6.34] in the case of rotating
machines, and it was found that transient torques and linear forces could
exceed the nominal torque of the motor, in some cases by even ten times (for
instance for fast self-operated repeated switching on and off of induction
motors). This is the effect of strong coupling of electromechanical and electro-
dynamic fields. There is much less information on this subject in the area of
windings of large transformers, end-windings of large electrical machines,
busbars etc.
In reference [6.31] a method was proposed for the generalized solution
of vibration and transient displacement of conductors taking into account
changes of inductances. The task consisted of solving the investigated system
of transformer windings or end-windings of electrical machines (Fig. 6.5) by
subdividing the winding into elementary bars, evaluating the field of elemen-
tary transient forces from the principle of virtual work (6.56)
Fx aWl]
= (-..!!!. , (aWl]
Fy = -..!!!. , Fz = [aWl]
__m
ax ;j=conslanl ay ;j=conslant az ;j=conslanl
(6.61)
and adding the elementary forces in such a way as to obtain the field of
components Fn Fy , Fz of electrodynamic forces in the form
_ 1 ~ ~ ..
F . - - LJ LJ l·h
aMjk(x;) (6.62)
I 2 j= k=
I I J ax;
where i = x, y, z and Mij is self- or mutual inductance of particular elemen-
tary conductors.
JCkI~
-·-·_·_·~·-6-
(a) (b)
Fig. 6.5 Calculation of components Fx• Fy • Fz of the field of forces by the method of
virtual work: (a) transformer windings. (b) end-winding of electrical machines
For the elementary conductors we attribute corresponding masses m l , m2

per unit length of conductor, damping coefficients D xl , D x2 ' DYI> D Y2 due to
external and internal friction, the mechanical work of forces etc. and the
elasticity constants K xl , K x2 , K yl , KY2 of the material, construction stiffness,
spring action etc.
The process of vibration and its amplitudes for conductor pairs or parts of
windings can be calculated using the Euler-Lagrange equation (6.23), taking
into account the mutual interaction of electromagnetic and mechanical
coupled fields during the impulse of short circuit.
In the case of two bars (Fig. 6.6) we obtain the system of six electro-
mechanical equations for electrical qk = Qt, Q2 and mechanical qk = Xl' X2'
YI' Y2 coordinates and generalized velocities Qk> Le. <21 = iI' <22 = i 2, XI = Vxl'
X2 = Vx 2' YI = Vyl' Y2 = Vy 2. The inductances Mij(xy) are calculated from
the classical formulae [6.11; 6.32], and then a system of partial differential
equations is formulated [6.33] which is solved with the help of a computer
library program, for example using an interactive program GOSPEL [6.33].
This package solves a system of nonlinear partial differential equations with
variables XI> X2' ••• , X n • The program enables not only interactive analysis of
the system, but also selection of the most efficient algorithm from several
numerical methods available. In addition these methods can be changed
during the analysis by the computer, to select the most suitable method at any
one time.
In reference [6.32] the principle of calculation of elementary electrodynamic
forces on the basis of equations (6.61) and (6.62) is given. In this approach
the method of 'mean geometrical distances' for the calculation of self- and
mutual inductances Mij of the conductors has been applied.
252 Coupled Fields
Oxl
)-~I--' A---tY,
Fig. 6.6 Elementary pair of busbars for calculating the field of forces and winding
vibration
6.2.9 Maxwell's stress tensor

Maxwell's stress tensor T is used for the expression of volume forces (6.50)
and (6.52) inside a body with the help of field and forces on the surface of
the body. Examples are the formulae (6.58) and (6.60).
For linear media with charge p = div D and current density J = pv =
curlH - EiJEliJt, the resultant surface force is
F = J fLdV= ~TodS (6.63)

v s
where the Lorentz force density [6.8] is
fL = pE + J x B = S[EdiVE + B x ~~J - J.1H x curlH (6.64)
From Gauss's theorem and eqn (6.63) it follows at the same time that
fL = divTo (6.65)
where TO is the vector of surface force, which is the projection of Maxwell's

tensor on a normal to the body surface and n is the normal vector.
Generally, Maxwell's stress tensor at S = constant and J.1 = constant is
Example
The force F acting on a surface S of the stator or rotor of an electrical machine
is
F= HTdS
S
= H~ [(B'O)B - ~2 IB I20] dS
s J.1o
= (J.10Il H;, ; Hf dSJ 0 + (J.10 II HnHtJ dSt = Fno + Ftt
(6.66a)
It can be seen that two forces Fn and F t act on each metal element. Nor-
mally motor cores are made of laminated iron and are not highly saturated.
Therefore one can take H t =0, hence force Ft =0 and on the iron core there
only acts the normal force Fn = 0.5Hs BH dS.
If however the core is made of solid iron, a permanent magnet of low per-
meability, or if it is covered with a copper screen and the field is alternating,
equivalent permeability J.1 is low and the values H n and ~ are comparable. In
this case both forces F n and F t exist and should be considered, especially at
impulse excitation of switchgears or fast relays.
From eqn (6.66a) it can easily be seen that the modulus of the surface force
density of the magnetic field satisfies eqn (6.60) and vector H divides the angle
between 0 and T in half (Fig. 6.7a). In this way the electromagnetic volume
forces can be transferred to the surface.
These forces in general do not coincide with the direction of the field H or
the vector 0 normal to a surface (Fig. 6.7). Their behaviour can be compared
to an elastic belt pull [6.11].
Depending on the direction of H the force may be stretching the con-
ductor in the direction of the field (Fig. 6.7b), may be compressive per-
pendicular to the field (Fig. 6.7c), or may even act to move the conductor
as in Fig. 6.7d.
Depth of penetration of magnetic pressure

In the case of the action of the magnetic field impulse H = H,(O, t) along
a metal surface (Fig. 6.7c) we introduce to eqn (6.50) the current density
[6.11]
J = _ aH, (6.67)
x az
(a) (b) (c) (d)
Fig. 6.7 Maxwell's stress tensor ITI = 0.5 pH2 according to the position of field H on the
body surface
254 Coupled Fields
Integrating this force with respect to the z-axis we find the hydrodynamic
pressure on the depth z brought about by the coupling of the magnetic field
with the incompressible body
(6.68)
where
(6.69)
is the pressure of the magnetic field on the surface. From eqn (6.68) it follows
that a pressure drop in time and space inside the metal is
ap aHo aH ap aH
- = IlHo- -IlH- and - = -IlH-
at at at az az
This means that the full magnetic pressure decays faster in time than the
external field and acts on a smaller depth than that of the field penetration.
In the case when the conducting body is compressible an impulse of
hydrodynamic pressure penetrates the metal with finite speed. According to
Knoepfel ([6.11], p. 125) if the magnetic field strength is of the order of
megaoersteds (several hundreds Am-I), and the pulse rise in time is of the
order of microseconds, a shock wave develops (one or several). In this case
it is necessary to consider that a diffusion effect of the field is linked with a
pressure distribution in the conductor. This relatively complex dynamic
problem is resolved with the help of a number of hydrodynamic equations,
including Maxwell's equations.
6.3 MHD
Magnetic hydrodynamics (magnetohydronamics), MHD, has a broad range of
applications: from astrophysics and thermonuclear reactions to typical indus-
trial equipment such as in transport, melting and mixing of liquid metals and
electrolytes, plasma motion in the electric arc of electrical apparatus etc. Here,
a complicated interaction between a conducting liquid media and an electro-
magnetic field occurs.
These phenomena are generally described by the system of equations created
from Maxwell's and Navier-Stokes' equations, which extend the Navier-
Stokes equation with the additional component
(6.70)
This represents the volume density of the Lorentz force (6.64), driving the
conducting media to motion with a velocity v.
The Navier-Stokes equation which describes the heat exchange by convec-
tion between the fluid and solid body surface, as well as the vector field of
fluid velocity v is usually presented in the form
MHD 255
av 1 v
-;- + (v grad) v = --gradp + vfi.v + -grad(divv) + P (6.71)
vt Pm 3
where:
Pm = Pm(X, t) fluid density,
p =p(X, t) hydrodynamic pressure,
V coefficient of kinematic viscosity of incompressible fluid,
vfi.v density of force of internal friction,
P force acting on mass unit.
Here the vector field of forces P can be not only the field of Lorentz forces
(6.70), but also the gravitational field etc.
The Navier-Stokes equation (6.71), together with Maxwell's equation,
create among other things a basis for the analysis of induction pumps and
liquid metal mixers [6.29]. It is difficult to obtain the solution of the system
of coupled equations of magnetohydrodynamics, therefore various simplifica-
tions are used, especially in analytical solutions. There are a number of analy-
tical solutions in this field, for example Sajdak [6.28; 6.29] and others. The
solution of fields equations (J, B) is first carried out at the assumed value v = 0
or v = constant, then by using the Navier-Stokes equations the electrodynamic
force densities are resolved. If the speed obtained is much different from that
assumed, the calculation is repeated.
In reference [6.28] the calculation of volume force density (6.70), where
J
aA
= 0 ( -- + v X curl A) (6.72)
at
is carried out from the equation for a medium moving with velocity v
fi.A - J.10 (v x curl A) = J.1o-

aA (6.73)
at
In induction pumps and mixers a cylindrical model is normally used, which
allows a significant simplification of calculations of current density, forces,
power and efficiency of the system. The accuracy of the solution depends
mainly on the degree of simplification of the geometry of the system and on
its Reynolds number (6.74). At small Reynolds numbers R M <.1 (e.g., after
[6.28], at v=0.5ms- 1 and R=0.05m, for liquid Al R M =0.16 and for Zn
R M = 0.085) one can assume a model with v = O. For a large capacity and
high metal speed and low frequencies (R m large) ignoring the motion of the
metal can lead to significant errors [6.28]. Among the several simplified solu-
tions of electric heating problems which have been published, using Navier-
Stokes equation, Wieczorek (1st Conference on Mathematical Methods in
Electrothermics, MMET; Wisla, Poland, 1986) proposed an analysis of induc-
tion heating processes of metals, which comprises interdependent electro-
magnetic, thermal and thermoelastic phenomena while neglecting the dynamic
effects and the effect of coupling of the fields of stresses and temperature.
In this case the task reduces to the consecutive, separated solution (Fig. 6.8)
256 Coupled Fields
Initial Electromagnetic Electrical

asumptions problems parameters
A,B," of inductor
Processing Maxwells equations S,E,J
requirements Equivalent
scheme
p f
v v
U ~=R+jX
Thermal problems Thermoelastic 11 cost/>

e
Kirchhoff-Fourier problems
equations Duhamel-Neumann
equations
T(r, t) liT
u CT
e
1j
General method of design of the induction heaters
Dimensions N,I,U,P,w,grad T,th, compensation, symmetri-
'--
zation, cooling, insulation, permitted superficial power,
mechanical parameters, structure
Fig. 6.8 Flow chart of the complex analysis and design of induction heaters of metals
(Wieczorek, MMET, Wisla 1986, Poland): where N, I, U, P, w, th are the number of turns,
current, active power, angular frequency and heating time respectively, Pv is the volume
active power density in metal charge, f v is the volume density of electrodynamic forces in
the charge; Z is the system impedance; 'leo 'IT are the electrical and thermal efficiency of the
system respectively; T(r, t) is the non-stationary, source temperature field in the charge; u, U,
e are the vectors of dispacements and stress, and deformation tensors in the charge; and A,
B, H, S, E, J are the vectors of the fields explained in text
of equations of electrodynamics (Maxwell), thermal conductivity (Kirchhoff-

Fourier) and quasi-static thermoelastic (Duhamel-Neumann).
In electroheating problems and in the heating of electrical machines and
devices, the structures and processes are complex, and so coupled fields are
still often resolved using an analytical-numerical method rather than methods
based on volume discretization. In such cases when coupling of both fields
is weaker and computational 'decoupling' of the electromagnetic field from
the thermal one is possible, the finite-element method is often successfully
applied, e.g. in the works of Krsteva et al., Schultze and Andree, Napieralska-
Juszczak and Dems and others (see papers 11-1,11-3 and 11-6, respectively, of
the MMET Conference, Wisla, Poland, 1993).
According to Voldek ([6.35], p.29) it is impossible to obtain full solution
of the magnetohydrodynamic equations in MHD induction a.c. machines, and
simplifications in this case are absolutely necessary. Since liquid metals and
electrolytes can be considered as incompressible, and as their temperature
along the MHD machine is practically invariable, it is sufficient to add to eqn
(6.71) the equation of continuity of incompressible liquids
divv =0
and consider, after Voldek, that the liquid secondary circuit acts as a solid
body, Le. as a secondary part of the linear induction motor.
MHD 257
Magnetic Reynolds number

As a measure of the degree of coupling of electromechanical and electro-
magnetic primary and secondary fields the magnetic Reynolds number
L v
R m = vl!oL = - = - (6.74)
L e Ve
can be used. It is a dimensionless parameter, in which v is the velocity of liquid
flow or motion of the conducting body, I! is the magnetic permeability, is °
the electric conductivity, L is the linear dimension of the field of flow (e.g.
radius of channel with the liquid metal [6.30]) and
1
L e = - or (6.75)
VJ.10
L e is the characteristic length, over which the conducting medium is subjected

to the magnetic field and Ve is the characteristic velocity with which the
magnetic field is propagated ([6.30], p. 135) in a conductor
If R M:> 1 (i.e. L :> L e or V:> ve ) the magnetic field moves jointly with the
conductor and the entrainment of the field or its extension by the conduc-
tor will prevail over its diffusion (freezing). Such effects takes place very
strongly in geophysical and astrophysical phenomena. On the other hand,
when R m < 1 (i.e. L < L e or v < ve ) the motion of the conductor does not
observably influence the magnetic field. This occurs, for example, in labora-
tory experiments with the flow of mercury, sodium or salt water.
In electrical induction machines with a relative speed of the field with
respect to the rotor
son sO) 1tD sroD
V2=-=--=--
1t 1t 2p 2p
the magnetic Reynolds number (6.74) can be expressed as
't 2't DL
R M = vl!oL = srol!o-L = 2k -L =-2 (6.76)
1t 1t po
where
nl - n ~
(s = -n-t-' 0) = 21tf J (6.77)
which is equivalent to the depth of penetration of the field into the solid
rotor.
In electrical machines, the Reynolds number is large only if the slip s
is sufficiently large. This is due to high copper conductivity (OCu,200C =
58 X 106 S m -I) and high iron permeability (J.1, = 2000-4000).
258 Coupled Fields
o
rlf:-
L
Fig. 6.9 Normal flux density distribution in a gap of the induction motor at different slips
s [6.6}:
- - LIM with finite armature length;
--- rotating and LIM with armature of infinite length
At a synchronous speed of induction machines (s = 0) R m = 0 and the field

penetrates the rotor without any obstruction. With the rotor at standstill
(s = 1), on the other hand, R M reaches its highest value (at the smallest depth
of penetration 0 of the field) and then the stator field is strongly deformed
by the reaction of the rotor currents. In the intermediate states (0 < s < l) the
rotating field entrains the rotor and is deformed at the same time (Fig. 6.9).
Smaller oscillations in Fig. 6.9 are caused by the influence of slots.
In rotating and linear induction motors of infinite length (L = 00; dashed
lines in Fig. 6.9) the change of slip from s == 0 to s = 1 brings about over
twofold suppression of flux density in the air-gap. In linear motors with a
short armature (L = 0.123 m; continuous lines in Fig. 6.9), diffusion damping
of the wave front under the moving armature and an electromagnetic tail on
the abandoned armature edge occurs. This is due to edge armature effects,
even at s == O.
Similarly a twist of the linear motor axis by an angle a with respect to the
direction of motion of the secondary brings about significant field deforma-
tion (Fig. 6.10), which effects a decrease in the drawing force, causing addi-
tional power losses and a force realignmenting of the twisted armature.
The field deformations can be considered as a decayed imperfect 'freezing'
of the armature field on the conducting surface of the secondary, which
according to Lenzs' rule acts towards maintaining the initial state. In this
case in the Reynolds number (6.76), instead of s = (VI - V)/VI' as for motor
MHD 259
1 v-const
lI.32 0.=0
Fig. 6.10 Deformation of flux density distribution in the air-gap of LIM with the twist of
the armature axis with respect to a forced secondary speed v = O.9v1 {6.17j
with L = 00, we put Sf = V!VI == 1 to 0.5, which gives the large value of
RM •
In linear induction motors the Reynolds number is expressed [6.22] in the
form
(6.78)
where d is the thickness of the conducting surface of the secondary.
In these motors the Reynolds number changes from R M = 0.265 for small
motors with vl=3ms- l , oAl=35x106 Sm- J, d=2mm, through R M =8
for average motors with VI = 100 km h- I , <Jcu = 57 X 106 S m- I , d = 4 mm, up
to R M = 200 for high speed trains with VI = 500 km h -I, <Jcu = 57 X 106 S m- I
and d=20mm.
The magnetic Reynolds number R M , as a discriminant of coupled electro-
mechanical fields, can be considered as a quantity factor of Lenz's rule. This
rule reveals not only the inertial field deformation, but also the repulsive
action of the eddy current reaction and in the levitation of linear motors
during fast motion.
According to reference [6.22] the normal force F n (attraction and repul-
i-r
sion) and drawing force F i of a single-sided linear motor (Fig. 6.11) moving
with a slip s can be expressed
s
t [~J
(6.79)
Fn =_
21l'ob VI W 2 ( J 1t 1t )'
Ch~Oi + RMssh~Oi
(6.80)
where:
J.1o = 41t x 1O- 7 Hm- 1
260 Coupled Fields
F,.
~ ~.2' 'RIlATRRE
F; V.VI~-t::: Al SECONDARY
·l\\\\\,(\\\\~ fe
Fig. 6.11 Single-sided linear induction motor
EI emf in primary winding,

WI quantity proportional to number of turns per pole pair,
VI = 2ft synchronous speed,
t' pole pitch
b magnetically active width of motor,
0; = ogapkC ideal gap,
kc Carter's coefficient.
From eqns (6.79) and (6.80) it foll0ws that at the synchronous speed (s = 0)
only the pulling force F n exists and there is no drawing force F;. At s > l/R M
the levitation (repulsive) force exceeds the attractive force. Since 0 < s < 1,
the levitation effect will not occur for small low-speed motors having
R M < 1. Levitation only appears during braking conditions (s> 1) of small
motors, or in large motors with large values of R M , of the order of 10 to 25
and more.
6.4 NON-HOMOGENOUS MEDIA

In the electrodynamics of superconductors ([6.30], p.117) the London's
equations are
a(AJs )
curl AJs = -B and at =E (6.81)
where Js is the density of the superconducting current, A = m/(e 2 ns )' the

constant parameter of the superconductor; m, e, ns are the mass, charge and
density respectively of the superconductive electrons and become a neces-
sary addendum to Maxwell's equations. London's equations (6.81) take into
account the structural changes of a superconductor under the influence of the
external effects of field and temperature and help to explain the Maissner and
other effects.
Another quite independent structural correction derives from other assump-
tions by Rawa [6.25] with respect to all kinds of non-homogeneous media in
which physical parameters of the medium are continuous functions of the
coordinate vector x, Le.
a = a(x) = a, E = E(X) = E, ~ = ~(x) =1I (6.82)
The work of Rawa has led him to the conclusion that the classical formulations
of Maxwell's equations
Coupled Magnetic and Thermal Fields 261
aD aD
curiO = J +~ and curiE = - - (6.83)
vt at
with the condition diu D = 0 do not describe sufficiently the electromagnetic
field in non-homogeneous media. From reference [6.25] these equations are
only satisfied if we assume diu D = p, where p is not defined solely as the
density of free charge introduced externally to the investigated region. It is
instead a quantity which characterizes the variable concentration of free
charge (e.g. electrons), created as a result of medium non-homogenity. In
order to eliminate this complication Rawa has proposed complementing
Maxwell's equations (6.83) with the additional equation
ap a E
- + - P = aE . grad In - (6.84)
at E a
Only in the special case, when Va == constant, are eqns (6.83) satisfied without
correction, this occurs in homogeneous media or in special cases of non-
homogenity. This hypothesis, in order to be confirmed, needs further inves-
tigations connected with the electron conductivity (0), dielectric irregularities
of solid bodies, (&), and the dynamics of charge motion (-ap/at = J).
6.S MAGNETOSTRICTION
Magnetostriction is the name for the complex phenomena connected with
the change of crystals and body dimensions under the influence of magnetiza-
tion (Fig. 6.12). This is caused by the change in separation between atoms of
the crystal lattice. Magnetostriction, which is characterized by unit elongation
E = Ii. III can be positive or negative. It is one of the causes of noise in electrical
machines and power transformers, and is used in ultrasonic generators.
6.6 COUPLED MAGNETIC AND THERMAL FIELDS

6.6.1 Magnetoelasticity
A typical example of coupled fields is the complex phenomena called magne-
toelasticity. This is a relatively new branch of science, which is devoted to
investigating the creation of additional electromagnetic fields as a result of
mechanical and thermal disturbances in a body placed in a strong magnetic
field. These additional field effects depend on the value of stresses, and the
speed of propagation of elastic and electromagnetic waves ([6.21], p.7).
The essence of the problem is the fact that if a body is placed in a strong
external magnetic field HO, the mechanical impact or thermal shock gene-
rate within it a field of deformation coupled with an electromagnetic field
([6.21], p. U8), which is characterized by small fluctuations b(x, t) and e(x, t)
which are excited in the body by mechanical or thermal effects. From the
moment of activation (t = 0) the resultant field at the moment t> 0 has the
form
262 Coupled Fields
£=tJl/l
ItfO-6
6
5
-I
3
2
1
O~~~~~~~
-1
-2
Fig. 6.12 Magnetostriction curves of transformer steel:

I. hot rolled after Knowlton,
2. cold rolled Armco,
3. cold rolled not annealed after BBC,
3 '. as 3 but annealed,
4. cold rolled not annealed
H(x, t) = HO + h(x, t), E = e(x, t) (6.85)

where x is a point of body moving with a velocity v in an external electric or
magnetic field.
At small velocities v of displacement of deformed body the constitutive
equations can be expressed ([6.21], p. 120) in the form
D = EE + av x H
B = IlH - au v x H au = Ell - Eollo
J = (E + v x B) + pv
(J (6.86)
In the theory of magnetoelasticity it is assumed that the only mechanical effect
in the electromagnetic field is Lorentz's force ([6.21], p. 120) with density
(6.64) which is developed from the equations of motion (6.87).
Using the principle of conservation of momentum and after several math-
ematical transformations, Nowacki [6.21] has developed a system of displace-
mental equations of motion of an isotropic body without initial stresses
Il LdU + (J"L + ilL) graddivu + X

a2u
+ fL = Pm at 2 (6.87)
where ilL' I..L are the Lame constants concerning an adiabatic state, u is the
vector of deformations (displacements Ui) with tensor of deformations Eij and
tensor of rotation wi) ([6.21], p. 15) which are described with the help of
equations
(6.88)
X is the vector of mass forces in an external load causing body motion; mass
forces are a function of position x and time t; the power of these forces is
equal to Iv X; u; d V; fL is the density of the Lorentz force (6.64) and Pm is the
density of the medium.
The complete set of differential equations of magnetoelasticity, apart from
eqn (6.87), encompasses further equations of electrodynamics (6.83), diver-
gence equations (6.24) and constitutive equations (6.86).
The boundary conditions for these equations ([6.21], p. 123) are
[E + v x BL = 0, DL = Jms -
[H - v x Psvm (6.89a, b)
[OJ; + 1); + g;vj]nj = 0 (6.90)
where the left-hand side of eqn (6.89b) represents the difference between the
tangential components on the surface both inside and outside the body; Jms'
Ps are surface values and OJ; is the symmetric stress tensor, connected with
external surface load p(x, t) by the equation [6.21]
p; = 0jinj (6.91)
where nj is the component of n normal to the surface S of region V
1); = E;Dj + H;Bj - HEkDk + HkBk)oji (6.92)
is the Maxwell's stress tensor (6.66); OJ; is the Kronecker's delta; and
g; = (D x B); (6.93)
Electromagnetic-pulse forming
The immediate application of magnetoelasticity in electrical engineering is in
electrodynamic forces in transient states, and especially in the magnetic-pulse
forming of metal parts. These are mostly cylindrical or flat elements [6.3].
Maxwell stresses (6.66) acting on a conducting surface are produced by dis-
charging a battery capacitor through an excitation coil. The coil produces a
damped out, oscillating impulse with a magnetic field amplitude of the order
of 20-50 tesla. Stress impulses produced in this way according to reference
[6.3] reach 109 N m -2 in about a dozen microseconds.
Krajewski [6.13] has resolved the system of magnetoelasticity equations
(6.87) for infinitely long cylinders in which the field is excited with the help of
• a cylindrical coil or
• a rod inductor.
Considering an axisymmetric system, with assumptions of linearity,

isotropy and homogeneity of the medium, Krajewski has obtained the
following system of equations in cylindrical coordinates
a 2H 1 aH 1 a aH
- +- (1 - olJ.vr) - - olJ.-H- (rv) - OIJ.- = 0 (6.94a)
ar 2 r ar r ar at
264 Coupled Fields
(6.94b)
(6.95)
(6.96a)
FL = -jlH [-aH + -1 H J (6.96b)

ar r
V=-
au (6.97)
af
Equations (6.94a) and (6.96a) describe the circumferential excitation currents
whereas eqns (6.94b) and (6.96b) describe the axial current, where H = Hz
for the circumferential excitation (a) and H = H", for the axial excitation
(b)
FL = FLF, V = v" u = u,
where E is the modulus of elasticity, and v is Poisson's coefficient. Equations
(6.94) to (6.97) have been resolved by Krajewski [6.13], using the following
iteration method.
The solution of eqns (6.94a, b) for a stationary medium (v = 0) provides the
first step of approximation Hj(r, f), J.(r, f) for the electromagnetic field dis-
tribution. The Lorentz force, calculated from these distributions, was used to
evaluate the first approximation of the displacement field u.(r, f) and the
field of velocity vj(r, f) with the help of equation of motion (6.95). Using the
values u. and v., by applying the generalized Ohm's law (6.86), a second
approximation of current density
(6.98)
is obtained.
A second approximation of the magnetic field intensity H can be found
using Ampere's law:
(1) for a system with circumferential currents, when the cylinder is placed
inside a coil
'2
H 2 (r,f} = J J2 (p,f}dp + Ae{t} (6.99)
where Ae(t) is the specific electric loading of excitation current in Am- 1

(2) for a system with an axial current
,
H 2 (r,t) =~I pJ2 (p,t)dp (6.100)
'2
where rl and r2 are the internal and external radii of the cylinder.
Proceeding in this way Krajewski obtained consecutive approximations

of the electromagnetic field distribution. Since functions In(r, t) are obtained
in the form of discrete integrals (6.99) and (6.100) they are replaced by the
corresponding quadratures. The proposed procedure, using relatively simple
analytical solutions [6.13], obviates the need to find repeated solutions of the
relatively complicated equations (6.94a) and (6.94b). Next in reference [6.13]
a method was proposed to solve the hyperbolic equation of motion (6.95) with
the help of the Boundary Elementary Method (BEM). The method proposed
in reference [6.13] is useful for the analysis of electromechanical problems in
which the velocity of the medium is considerable, Le. where the magnetic
Reynolds number (6.74) is much higher than 1.
6.6.2 Magnetothermoelasticity
In electrical machines and electrical power equipment, especially of high
power, and in electrothermic equipment, the thermal effects resulting from
electromagnetic fields play an exceptionally important role. They very often
determine the reliability of these devices. Of lesser technical significance from
this point of view is magnetothermoelasticity. Nevertheless it will be men-
tioned now for completeness of the description.
Equations oj magnetothermoelasticity
Nowacki [6.21] and Parkus [6.24], in a similar way to eqns (6.40) and (6.41),
as a starting point to the analysis-6~ magnetothermoelasticity, have taken
balance equation of mechanical, electrodynamic and thermal energy, which
in space notation for a volume V with surface S reads
:tt(iPmVjVj+PmW+ ~JdV= t(XjVj+PmQI)dV
r
+ [PjVj - qjnj - (E x H)j + ~vjnjl dS
s
(6.101)
where:
21 -dd r PmVjvjdV increment in time of kinetic energy;
tv
1
d Pm WdV
d tv
increment in time of mechanical energy;
d 1~dV increment in time of internal electrodynamic energy;

dtv
266 Coupled Fields
Pm and v density and velocity of body;

W mechanical and electromagnetic internal energy
density per unit mass;
~ = 1(&£2 + ~eH2) electrodynamic energy density per unit volume;
JX;v;dV power of external mass forces X (see eqn 6.87), which

v cause motion of the body and are the functions of
position x and time t ([6.21], p. 13);
thermal power; QI is heat quantity generated per

unit mass and time;
power of external, surface (contact) forces which

cause motion of the body and are functions of posi-
tion x and time t;
Jq;n;dS flow of thermal flux q; through the surface S, where

s n; is the vector normal to the surface dS;
J(E x H);dS flow of energy of electromagnetic field;
J~v;n;dS electromagnetic energy flow caused by body motion

s in an external magnetic field;
q density of thermal flux vector;
(E X H); ith component of Poynting's vector.
For a body moving, or deforming, slowly with velocity v, Ohm's law may
be applied. It is written here in a modified form to account for body deforma-
tion (v x B) and heat flow under the influence of temperature e = T - To
J = a(E + v x B - 1tograd9) (6.102)
Considering now equations of mass Pm continuity and entropy 11 bal-

ance, Nowacki [6.21] derived a set of three equations of magnetothermo-
elasticity
A
U 1 ~J
- - l /t e -11 d"'
l V U =PmQI
--- (6.103)
[
K ko
where:
K = ko/Ce , ko > 1, Ce constants; at = alat;
11 = ~/ko entropy referred to mass unit;
~=a'To, a' =(3A,+2~')at, ~', A, Lame constants in generalized Hook's
equations referred to the adiabatic
state;
coefficient of linear thermal expan-
sion; ko > 0 a constant;
u vector of displacements at body
deformation;
Pm and QI density, heat generated per unit mass

per second.
(~ -13o,)H -13oo,2H = -13cur/(v x H) (6.104)
where 13 = ~o, 130 = ~e;

~'~u+ (A + ~')graddivu + x + ~[(curIH-eE) x H] = Pmu + 0' grad6
(6.105)
The equations of magnetothermoelasticity (6.103) to (6.105) illustrate the

complexity of coupled fields which have been discussed, in spite of the fact
that they still contain limitations due to the assumption of slow moving media
and simplifications in the form of ~ =constant, e =constant and P =o.
Thermoelasticity, as a part of elasticity theory plays an important role
in the investigation of stresses and strength of structural elements under
the influence of changes and non-uniform distributions of temperature. It
concerns particularly the end regions of rotors, rods of squirrel cages and
short-circuiting rings of large induction motors, winding barstocks of turbo-
generators etc. Uncontrolled dilation (linear L T = L o(1 + A~ n
or volume
VT = Vo(1 + a~n thermal expansion), can cause serious failures, espe-
cially with the assembly of parts of different coefficients of thermal expan-
sion, as for instance Acu = 16.5 X 1O- 6 K- I, AAI =24.1 X 1O- 6 K- I, Abronze =
17.5 X 10- 6 K- I and ASteel = 11.10 X 10- 6 K- I. For isotropic bodies a =3A.
6.6.3 Thermokinetics
In electrical power equipment it is not usually necessary to apply a full set of
coupled equations (6.71) and (6.101). Indeed, the difference in time constants
between electromagnetic, thermal and mechanical processes is such that they
may be studied independently, i.e they are weakly coupled. However, strong
coupling may be found in certain cases, for example in induction surface hard-
ening [6.30].
The fundamental equation of thermokinetics dealing with heat exchange
through contact is contained in the Fourier-Kirchhoff's law [6.10].
~6vt + v grad 6 =..!:.J...

cPm cPm
a;
+ _1_ [diV (A grad 6) +
vt
+ v grad P ] (6.106)
where:
6 = 6(x,Y, Z, t) temperature in K;
t time in s;
v = v(x,y, z, t) velocity of the medium in m s - I;
PI power of the heat source in W m -3;
=
P p(x,Y, Z, t) pressure in N m- 2;
c specific heat in J kg -1 K- I;
Pm material density in kg m- 3;
A = A.(x,Y, Z, t) thermal conductivity in W
268 Coupled Fields
in anisotropic media A is a symmetric tensor.

In the case of solid bodies (v = 0, p = constant) eqn (6.106) takes the form
of the heat conduction equation
ae
CPm = at = divAgrade + PI(x, t) (6.107)
The electrical conductivity 0 and magnetic permeability ~ of the body are

often temperature dependent. Thus a feed-back loop is created, where the
temperature field reacts, via 0 and ~, on the original electromagnetic field and
power loss which provided heat sources.
For thermally isotropic media (Ax = Ay = Ax = A), using the identity
div grade =v2e, we have
1 ae 1
V e---=
2
K at
--PI(x t)
A '
(6.108)
where K = A/(cPm) is the coefficient of equalization of temperatures, called

thermal diffusivity.
6.6.4 Thermometric method of power loss measurement

Equation (6.108) provides the basis for the thermometric method of measure-
ment of local power losses in W m -3 and magnetic field strength H ms on the
surface of the body. The power PI developed in a unit volume inside the
body can be expressed in terms of the temperature e as
ae [a 2
e a2
e a2
e]
PI = CPm at - Ax ax2 + A a 2 + Ax az2
y (6.109)
y
In the case where the sample is in the form of a thin sheet (Fig. 6.13a) in which,
in its whole thickness, a uniform power PI is developed, the thermal balance
equation (6.108) should include a term to account for heat dissipation to the
environment of ambient temperature eo
PI = CPm ae [a 2e a2e a2e]

at - Ax ax2 + Ay ay 2 + Ax az 2 + a' (e - eo) (6.110)
where a' is the resultant coefficient of heat dissipation to the surroundings

through convection and radiation in Wm- 2 K- I • If the body temperature is
measured sufficiently rapidly at the instant of field excitation (t = 0), so that
the temperature of the body may be considered uniform, i.e.
a2e a2e a2e
v2e = Ax ax 2 + Ay ay 2 + Ax az 2 = 0 and e = eo (6.111)
then from eqns (6.109) and (6.110) we have
PI = CPm [~e]
vt 1=0
= CPm tan ao (6.112)
(a) (b)
Fig. 6.13 Measurement of unit power losses W m -3 with thermometric method

PI =CPm tan ao (/6.301. p. 406-416): (a) method of measurement. (b) reading of tan ao from
the curve
where a.o is the gradient of the initial rectilinear part of the heating curve
(Fig. 6.l3b). Performing as fast and accurate as possible a measurement of
the temperature with the help of built-in thermoelements (thermocouples) we
can evaluate from eqn (6.112) a power loss density distribution in structural
parts of electrical machines and devices [6.12) without endangering the equip-
ment by excessive heating during the test. A disadvantage of this method
is that it requires a sudden switching-on of full power. This can be difficult
for large objects (e.g. power transformers). Moreover the device needs to be
cooled down before subsequent measurements. This cooling can be avoided
if simultaneous automatic measurements are taken with many built-in thermo-
elements in a short space of time, as described in reference [6.12). Applications
and errors of the method are discussed in references [6.20) and [6.30).
The less heat outflow during the measurement, that is the more closely the
condition (6.111) is observed, the better is the accuracy of the thermometric
method. Best results will therefore be obtained in regions which have uniform
loss distribution, e.g. transformer cores. Largest errors occur when power loss
and temperature distribution are varying rapidly, which takes place if the skin-
effect is strong.
A typical example of such rapid changes in power and temperature
distributions is in solid steel walls (Fig. 6. 13a). In such solid plates the
gradient of the resulting power loss p\(z) is so large [6.30) that even during
the short time of measurement of the heating curve 9 = 9(t) there is con-
siderable heat outflow from the point of measurement to colder parts of
the metal. This significantly decreases the temperature on the surface and, at
the same time, the value of tan Uo and of the measured losses. This error
has been investigated in reference [6.20]. This is a typical example of coupled
fields described by eqn (6.108). If the analysis is limited to a plane wave, we
have
270 Coupled Fields
028 1 08 1
OZ2 - ~ ot = -"iP1(X, t) (6.113)
The right-hand side of eqn (6.113) represents eddy-current power loss cal-
culated from distribution of the alternating electromagnetic field ([6.30],
pp. 180, 189). The distribution of volume power loss density P1(z) can be
calculated analytically only with the assumption of constant iron permeability
IJ., as
2
P1(Z) = -.!-.IJ (Z)\2= 2k ch2k (d - z) + cos2k(d - z) {ro lJ.lHms l
20 m ch 2kd - cos 2kd \J 20 2
(6.114)
For strong skin effect we can assume d-+ 00 and eqn (6.114) reduces to
[6.30]
(6. 114a)
In the case of ferromagnetic materials with strong nonlinearity IJ. = IJ.(H) it

is necessary [6.20] to apply a numerical method of solution of Maxwell's
equations (6.83) with nonlinear magnetic permeability
(6.115)
These equations have been solved numerically using a finite-difference method

by Zakrzewski [6.36] for plane polarized waves. The magnetization curve
B = f(H) has been automatically interpolated for every instantaneous value
H(z) at a depth z from the steel surface. The program has been applied in
reference [6.20] to the right-hand side of eqn (6.113). After harmonic analysis
of the non-sinusoidal dependence Em = Em(z, t) the resultant rms value for
each nth elementary layer has been defined by
1 m
E,ms, n = -';;'2 l: E~mn
'V1. v =1
(6.116)
where E vmn are amplitudes of corresponding time harmonics En(t) in the nth
layer. Only harmonics with amplitudes E vn ~ 0.01 Ev=l,n, where v = 1, 2,
3, ... , m have been considered. After introducing eqn (6.116) into (6.113)
the distribution of volume power loss density P1(z), necessary for further
numerical solution of eqn (6.113) has been obtained (Fig. 6.14). The solution
for 8 of the full equation (6.113) has been compared with an idealized case
of 8id , for any adiabatic (08/oz = 0) process. For such a case
08 id _ _ 1_ p
- I (6.117)
ot CPm
From the comparison of both solutions at the same value of !it and H ms
the correction coefficient k l has been determined to account for heat outflow
Fig. 6.14 Volume power density PIfz) distribution at JI =JI(H) inside solid steel wall for
different H ms values on its surface:
1-5: 10.0, 8.82, 7.0, 6.17, 0.4 kA m- J
1a: linear model exp(- 2kz) at H ms = 10 kA m - J, Jl r = 150 = constant, k = 450 m - J [6.20}
from the point where the thermoelement is attached away inside the steel
as
k -
1 -
I I -I a I
Plreal
Plmeas 1=61 -
aid
1=61
(6.118)
In this way the actual (real) loss PI real is k l times greater than the measured
P 1measloss after time t:.t. As t:.t-+O, Plmeas-+Plreal and k l -+ 1 (Fig. 6.15). The
stronger the excitation field the more uniform is the power distribution at a
depth d inside the plate and the smaller is the correction k l • The curves in
Fig. 6.15 refer to plates placed in still air.
Using a similar method, reference [6.20] presented a method and curves
which allow the evaluation of H ms on the steel plate surface, on the basis of
kl~i ~~
o~z~y.~
10b\10"~
10
8 8 / '1fPd.A/m
6
4
6
4
l:~~~
1/ / ......
I::'~/
2 ~"
----------------- d 1 ---------------- d
12 16 20 24 28 mm) o <4 8 12 16 20 24 28 mm)
(a) (b)
Fig. 6.15 Coefficient k l = kdd) by which measured, at time t1t = 1 s, unit power loss
Ps = cPm tan ao should be multiplied: for (a) one-sided and (b) two-sided excitation [6.20}
272 Coupled Fields
body temperature, measured after 30 seconds. The method also allows a direct
measurement of the power loss density PI both from the side of the excita-
tion field, and from the opposite side of the plate, e.g. the outside of a trans-
former tank.
6.6.5 Local overheating criteria

Uncontrolled heating of structural parts of electrical machines, power trans-
formers, parts of switch-gear etc. affect their reliability and are among the
most hazardous.
In 1960 [6.30] the method of electromagnetic criterion Hms,perm was pro-
posed and has proved very useful for fast estimation and location of hot-spots,
as well as places endangered by excessive heating caused by eddy currents and
strong leakage fields. This method is especially convenient when combined
with the reluctance network numerical modelling of leakage fields (Figs 4.6,
4.15, 4.16, 4.21) in large power transformers.
As a criterion of local overheating, the author selected the permissible
tangential value of magnetic field strength Hms,perm on the surface of parts
under investigation, at which the body temperature t and its rise 9 = t - to do
not exceed the permitted values (as recommended, for example, by ClORE)
tperm and 9perm. Values of Hms,perm can vary for different materials, excitation
methods and heat dissipation ([6.30], pp.375-392).
The unit power loss density in steel walls of thickness d > A (A = 21to,
o = .y'[2/(ro~0)]), with one-sided excitation by an alternating uniform mag-
netic field of amplitude H ms (Fig. 6.16), in Wm- 2 can be expressed ([6.30],
p. 152) in the form
IHms l
p'V{(roilS]
2
PI = Q . (6.119)
20 2
Introducing this value to the equation of thermal balance (Fig. 6.16)

(6.120)
where n' = no(9/90 )0.25 is the coefficient of heat dissipation by convection

and radiation, and using the analytical approximation (Fig. 4.18) of the recal-
culated [6.30] magnetization curve in the form
JIl,H 2 = CtH + C2H2 (6.121)
where c i = 3to X to 2 Am-I and c2 = 7.9, we finally obtain ([6.30], pp.377-

381) the permitted field value (Fig. 4.21)
IHms,perml ~ 1962[vI(1 + 3.288 x to-Sc') - 1] (6.122)
where for steel at the temperature t
C
9~;m -2
, = no, ---o2f ~(200
- .192 J 172 + t
, 0=00----
90' Qp ro~o 172 + tperm 172 + tperm
-i
11\
R1- 0 pvl1t'H..
T 6
OJ'@)
z
t t to
·d
·1
Fig. 6.16 Thermal balance in steel wall, one-side excited and cooled
If necessary, the values Hms,pe,m can be calculated by the numerical solution

of nonlinear eqn (6.108). .
Examples of the application of these criteria are given in Chapter 4. If the
permitted values are found to be exceeded other materials can be used to
increase the limits [6.30], or changes of geometry, screening and/or better
cooling can be applied.
Effects of non-uniform field distribution

If the magnitude of the magnetic field strength H ms on the surface of a metal
wall exceeds the permitted value given by eqn (6.122) we can expect excessive
heating. In this case a more accurate calculation should be carried out and
additional precautions mentioned above may be needed.
In large electrical machines and power transformers the permissible value
Hms,pe,m could easily be exceeded locally. However, the non-uniform char-
acter of the power loss density distribution can significantly reduce this
hazard. Typical example are the cover plate and tank wall of a transformer,
and clamping plates and end-plates of electrical machines, where the heat is
carried along the plate from warmer to colder parts (Fig. 6.17).
The non-uniform power loss distribution can almost always be approxi-
mated analytically with the help of the exponential function in two-dimensions
(6.123)
Since the plate thickness d is usually much smaller than others dimensions,
and the thermal conductivity I.. is large, the changes of temperature in the
plate thickness can be neglected, and only the changes along the plate need
to be considered. The heat exchange can be expressed by the equation
274 Coupled Fields
f}max
1\
e / \ Ae-8ln
~ '~t
,/ ""
"
",
, "" - y
Fig. 6.17 Non-uniform loss density and temperature distribution in a steel plate {6.30j
_ ae(y, t) _ A. a2e(y, t)
CPm - at - ay2 + q (6.124)
where the symbols are the same as those in eqn (6.106) and q is heat quantity
per unit volume, which comprises power losses 2P\ / d as a positive source,
and heat carried away a'e as a negative heat source, i.e.
2Pt !1y 2a'e!1y 2Ae- By 2a'e
q = ddy - ddy = d d (6.125)
Neglecting transients, eqn (6.125) was resolved [6.30] for the thermal quasi-
stationary state and the following maximum stationary temperature rise emax
was obtained (Fig. 6.17)
em. = l.dB'2A- 2«'

[B
~[~) - I
] ~ a'X
A (6.126)
where
1
K=--B---~ 1 (6.127)
---,--;,....--,,-- + 1
~(2~J
The coefficient K (6.127) accounts for the heat carried away from the hot-
spot to a colder place. At B = 0 or A. =0 the factor K = 1. In practice ([6.30))
K is often 0.5 to 0.7 or even less.
The values Hms,perm given in Fig. 4.21 can be considered as a criterion of
overheating due to eddy currents, and the coefficient K (6.127) as the neces-
sary safety factor in such cases. By dividing by K we can make the criteria
mentioned above more accurate (Fig. 4.21)
~~
em 1
A
Fig. 6.18 Influence of iron core lamination on decrease of eddy current losses in a steel
wall placed in a 50 Hz field of two busbars {6.30j
H _ Hms,perm
mS,perm, K - K (6.128)
By decreasing the value of K through changing the parameters present in

eqn (6.127), we can increase the value of the permitted field Hms,perm and so
reduce the thermal hazard of the part under investigation. In a more accurate
analysis of eqn (6.124) the matter of steel nonlinearity should be considered,
because this nonlinearity may have an additional influence on design. In
reference [6.30], pp.385-392 criteria for other materials, methods of field
excitations and different cooling conditions have been collected.
Effects of the lamination of steel walls

Power loss in laminated iron cores is much smaller than in solid steel walls
(Fig. 6.18). This results from both the reduction of eddy currents owing
to the laminated core, and the higher resistivity of silicon iron. This has
been confirmed by measurements on models (Fig. 6.18) with dimensions
b = 198 mm, cS = 50 mm, t1 = 10 mm and number of turns N = 288 in one 'bar'
(winding). Power losses in a solid steel wall of thickness 12 mm (curve 1), after
recalculation [6.30] in the ratio P5. =P.(02/o1)O.9 have given curve 2', which
almost exactly coincides with the curve 2 of losses measured in a stack of
laminates of Epstein's apparatus, when placed at right-angles with respect to
the flux lines. The same laminations placed approximately along the field lines
(curve 3) give significantly smaller power losses, and when placed completely
along the field lines (curve 4) produce negligible losses.
For the development of the criterion Hms,perm for the local overheating on
iron core surface (transformers, machines) the models should first be based
on curve 3 (Fig. 6.18). Such investigations, initiated by Mankin and Morozov
[6.16] have been extended significantly by Kazmierski, Kozlowski and others
[6.9], Fig. 6.19. Full solution of this problem requires three-dimensional (3D)
276 Coupled Fields
100
80
60
2
40
Fig. 6.19 Influence of the subdivision of external packet into n narrower packets on the
maximum packet temperature:
1. with a magnetic clamping plate,
2. with non-magnetic plates [6.9J
analysis, which poses a considerable challenge for contemporary numerical

techniques.
The leakage flux of density Be (Fig. 4.14) which penetrates perpendicularly
the surface of sheets and clamping plates of the core, may cause local over-
heating, especially at the top and bottom edges of windings. Such effects
were investigated using full scale models and are reported in [6.9]. It has been
found that the normal component of flux density on the surface of the core
distributes itself inside the external packet of laminations according to the
relationship
Bdy) = BdO) (1 + ar) (6.129)
where y is the distance from the symmetry plane of the packet and Bd 0) is
the normal flux density at the centre.
The local temperature rise is approximately proportional to UcPci, where
cPc = JBe dy (Wb m) -1. It follows that the best way of reducing the temper-
ature is subdividing the external core packet into n narrower packets. This
was confirmed by computations (Fig. 6.19) made for a 250MV A, 220kV
autotransformer.
6.6.6 Induction heating

Coupled electromagnetic and thermal fields occur in many kinds of induc-
tion heating, e.g. heating in forging and annealing, melting, transportation
and batching of liquid metals. This coupling is usually stronger than in other
electrical machines and devices. Typical examples of interaction of an electro-
magnetic field with a thermal field include the penetration of energy into metal
for induction heating or skin hardening of solid steel (Fig. 6.20).
The electromagnetic field distribution in metals in the initial cold state
(CS3t ~3) and the final hot state (CS2t 1J.2) may be described by the same
equations
Fig. 6.20 The penetration of electromagnetic fields H, E, J into steel for induction heating:
(a) initial, cold state; (b) transition state with increasing depth d of layer with Curie
temperature T = Tc and (c) final, hot state ({6.30j, p. 300)
IH2ms l
2
H
m
= Hms e- az '• m
J = oE = f!..H e-az.
moms ,
p
1
= 'VffoolJ,}
l20
(6.130)
only if the thickness d (diameter) of the body is larger than the wave-length
A in metal [6.30J, where
2 2
A = 21t f_ _ = 21to, 0 = f_ _ (6.131)
'VOOIJ,O 'VOOIJ,O
and 0 is the skin depth. If the thicknesses d < A the internal reflected waves
should also be considered ([6.30J, pp. 178-194).
The transition from the initial (cold) state to the stationary (hot) state is
accompanied in steel by an almost 9O-times increase of the skin depth 0 owing
to the l000-2000-times decrease of IJ, and the 6-1O-times decrease of o. As
E 3 = E 2 on the interface between layers 2 and 3 (Fig. 6.20b), a sudden jump
of current density J3 = 03E3 takes place on the surface of region 3. This helps
maintain uniform heating of the external layer of steel 2 for surface hard~
ening, without a risk of burning the surface of the hardened element. This is
278 Coupled Fields
z c
L-2(J+h)
Fig. 6.21 Application of the method of analytical prolongation: (a) core with inductor,
(b) infinite sequence of fictitious inductors, (c) periodic flow distribution ({6.32j, p. 289)
obviously only a qualitative explanation. As a first approximation such calcu-

lations can be made in a similar way as for electromagnetic screens ([6.30],
pp. 176-194). If necessary, the nonlinear permeability J.1 of the cold solid steel
layer can be considered using the corresponding linearization coefficients
ap , aq ([6.30], pp. 139, 318-328).
Calculations oj inductors using the method oj analytical prolongation

Induction heating is often applied to cylindrical metal charges by an induc-
tor of finite length (Fig. 6.21). Such a system is often solved with the help
of the so-called method of analytical prolongation, which uses a Fourier series
([6.30], pp.207-211, 333-345). This method gives in general good results
however the problem now arises in selecting a suitable distance 2h between
the fictitious inductors of an infinite sequence of mirror reflections. The
greater the distance 2h, the smaller the interference from the nearest fictitious
inductors. At the same time however, account needs to be taken of the higher
harmonics on summation of the Fourier series. For this, a small distance 2h
should be selected, but not too small to avoid interference between adjacent
inductors.
The optimal distance 2h can be evaluated as follows. The field in the air
gap of the inductor for the magnetic scalar potential Vm can be calculated
from Laplace's equation V 2 Vm = O. After solving Laplace's equation we can
find the components of the magnetic field strength H = - grad Vm , for a
given specific electric loading a = a(z), which in the case of cylindrical induc-
tors (Fig. 6.21) can be expanded into the Fourier series as
4a ~ 1. mel mez
a(z) = -; n~1 ;; SIn 2 (J + h) cos 2 (J + h) (6.132)
When studying the edge effects of the inductor we can use the approximate
formula ([6.32], p. 131) for radial flux density Bmr (Fig. 6.22)
8mi'
21 2 21
r-------
I r-r..,.-,,....,..T'"T"-'
,,/
"
Fig. 6.22 Evaluation of the optimal distance 2h o of the fictitious inductors ({6.30], p. 259)
(6.133)
where B mS is the maximum flux density in the gap 0 on the edge of the
inductor.
If we assume the smallest permissible value of the relative flux density
bo =BmO/Bms in the centre of the armature, we can evaluate the correspond-
ing optimum distance
2ho = ~1t cS [~bo - IJ (6.134)
which for bo = 0.05 gives 2ho = 240. More generally the condition may be
defined as 2ho ~ O.
Equation (6.134) refers to stationary inductors. If the inductor moves along
the z-axis (as is the case in linear motors or induction heaters using movement),
the reaction of eddy currents in the secondary circuit causes a weakening of
the field on the incoming edge of the core, and increase on the edge moving
away. The latter decays slowly in air and creates a so-called 'magnetic tail'
(Fig. 6.9). In this case the optimum distance 2ho significantly changes and
depends on the speed of the core (Fig. 6.23a). At constant speed it is a function
of the coefficient bs = Bm(x = 2ho)/B ms . The values can be evaluated from
graphs shown in Fig. 6.23. In order to avoid errors, the smallest value of
bs should always be used. In practice satisfactory accuracy is obtained by
assuming bs ~ 0.02.
Numerical methods for induction heating

The most popular methods nowadays for field computation are the finite-
element, finite-difference and boundary-element methods. However, in the
electrothermic industry these methods are used rather reluctantly [6.14],
because simple models often give quite satisfactory results.
280 Coupled Fields
J!!sL
t' 6
~
4 s-O,022 _
f- 3A 5 r--
bs-uos )
f- 3A
3 4
/
V 3
2
V
V
2 \
1 -" \
~
~ 1
~~
v, ~
o 0LS 1,0 o o.s 1,0
Q b
Fig. 6.23 Dependence of the optimum distance 2holr of fictitious armatures in a linear
induction motor SL-5-270: (a) on speed, (b) on assumed flux density bs
One of the oldest finite-difference formulations, still in use, for a heating

problem of a steel cylinder, uses two closely coupled differential equations
1rdr dHmJ
d [ r - - -uH
--
dr
2 =0
m
(6.135)
P c aT
m at
=!r ~ (rA aT] + q
ar at (6.136)
with a boundary condition on the surface
(6.137)
where:
0= o(r,9)
T temperature (K);
Q power loss density (W m- 3 );
Pm = Pm(T) mass density;
c=c(T) specific heat;
A=A(T) thermal conductivity;
s and a refer to surface and ambient temperature indices respectively;
uc coefficient of heat dissipation by convection;
Cs heat dissipated by radiation;
Qs constant surface losses (W m -2).
Levitation 281
Using simple transformations, the above solution may be extended to cross-

sections other than circular [6.14J.
In references [6.26; 6.27J the induction mixing of liquid metal using finite
differences has been solved. In this case the coupling between electromagnetic
and flow equations through the field v x B is weak and equations can be solved
independently.
For electroheat eddy-current or liquid metal mixing problems, reference
[6.2J reports a successful application of the volume integral equation. This
method is a very convenient approach to the solution of coupled axisymme-
trical, mechano-thermo-electromagnetic problems [6.14J. For this application,
the boundary element method, on the other hand, is seen as less useful.
Simulation of induction heating using finite elements, in contrast to the
problems of induction mixing, is shown to be very appropriate according to
reference [6.14J. However the final matrix equations for finding the value H m
in internal nodes
( [S (p )] + jro [ T][ Il]} [H] = (b J (6.138)
where P = p(T) and Il = Il(Ho, T) can be uncoupled from the thermal
equations
aT = -a (aT] a ( aT]
PmC-
at ax A. -ax + -ay A. -ay + w (6.139)
if the ferromagnetic conductor is in a non-magnetic state, Le. above the Curie

temperature (Fig. 6.20). In eqn (6.139) w is the internal heat source in
Wm.
-3
A comprehensive analysis of the state of the art of induction heaters and

their methods of computation has been made by Skoczkowski, cited in refer-
ence [6.33J. From his study it follows that most of the models applied so
far are one-dimensional and linear, and use averaged values of material
parameters. The most popular methods used are separation of variables
(Fourier), Green functions, the Laplace and Hankel integral transformations,
Fourier series and integrals, integral equations and others. Skoczkowski has
also elaborated the simplified numerical solution of the system of coupled
M~well's and Fourier-Kirchhoffs equations for induction heating of steel
charges, e.g. steel tubes. Applications of finite elements to similar problems
of coupled fields for thermal anisotropy of a body are also presented in refer-
ence [6.18J.
6.7 LEVITATION
Levitation is a technical term used to describe all phenomena connected with
lifting or suspension of ferromagnetic bodies under the influence of magnetic
fields of permanent magnets or electromagnets, or under the influence of
induced currents in solid conducting bodies placed in alternating fields or as
a result of rapid movement with respect to the field. Currently this effect is
282 Coupled Fields
gap assumed
position x
L..------=---~x
Fig. 6.24 Levitation of an iron cramp of an electromagnet [6.5J
used on an industrial scale both for suspension of trains on magnetic cushions

[6.19] and in technical processes (melting of metals, welding technology, sepa-
ration of steel sheets etc.), in machine design (magnetic bearings), diagnostics
(e.g. free suspension of bodies in aerodynamic tunnels) etc. Gieras [6.5] has
investigated linear induction motors with levitating secondary parts, driven
and stabilized electrodynamically, using electromagnets with automatically
controlled air gaps, magnetic bearings etc. In experimental systems of electro-
magnetic levitation (Fig. 6.24) the stability of air-gap x is achieved with the
help of negative current feed-back, controlled by sensing position, linear speed
v and coil current i. The operation of the magnet is described by the system
of equations
ax
- = v,
av 1 .2 dL(x) d
mat=mg+2'~' u=dt[L(x)i)+Ri (6.140)
at
where
Loxo
L(x) = L", + -x
, L", = L(x -+ (0) (6.141)
Xo is the nominal air-gap.

When applied in the feedback state variables x, m and i affect the supply
voltage u of the electromagnet according to the relationship
(6.142)
where K, Kv ' 1<; are the feedback coefficients and C(xz ' i) is evaluated for the
stationary state (x = x z ' v = 0) and equals
C(xz ' i) = u + K;i (6.143)
Equations (6.140) to (6.143) are used among others for the evaluation of the
stability zone of the system.
The force of attraction of the electromagnet with an air-gap x (Fig. 6.24)
can be calculated from the formula
Levitation 283
Fig. 6.25 Levitation of secondary element of linear induction with transverse field
(6.144)
where:
Ss cross-section of the air gap;
SFt! cross-section of the iron core;
1Ft! length of path of magnetic flux in the core and magnet keeper;
1-1, relative magnetic permeability of the core depending on coil current i.
In the model of the linear motor with levitating secondary element (Fig.
6.25) the inductor (armature) of the motor is made up from two rows of cores
with mk cores in one row, where m is the number of phases, k = 1,2,3, ....
The aluminum block has inclined sides angled at 60°, which ensures the
highest levitaion force with side stabilization.
The boundary condition for Laplace's equation in air and Helmholtz
equation inside the walls of the secondary is the distribution of the normal
flux density component on the inductor surface, which with a three-phase
supply with a sequence R, -8, T, -R, 8, - T, ... has the form (Fig. 6.26):
= 0.5Bo L: L: bvbn[ej(ro:t-llvX) (1 + a +
00 00
Bz(x,y) V 2
+ a 2v + 1)
v=1 n=1
(6.145)
where:
Bo flux density at the centre of sym-
metry of the main column;
v = 1,3,5, . space harmonics in the x-direction;
n = 1,3,5, . space harmonics in the y-direction;
284 Coupled Fields
Fig. 6.26 Magnetic flux density distribution on the surface of the inductor
O):=s:O); O);=S;O); 0) = 27t!; Sv slip of vth harmonics;

~v = V7tlt; TIn = n7t/'t y ; t; t y (Fig. 6.26);
a = exp ~2;l bv ; bn coefficients of Fourier's series.
Indices' +' and' -' indicate direct and reverse components of the rotating field
respectively.
Gieras [6.5] has also given a solution of the equations for the field dis-
tribution. He recommends the application of Maxwell's stress tensor for the
calculation of forces acting on the unit surface of a secondary
!zvn = -0.51-10 Re[HzvnH;vnL=1l (6.146)
!zvn = 0.51-10 Re [0.5 HzvnHivn - 0.5 HxvnH;vn - 0.5 HyvnH;vnl z=Il'
(6.147)
where cS is the air-gap dimension.
The total forces acting on the secondary part, taking into account that the
field is excited by two rows of electromagnets, are
Fx = 2L ~ J!xvn dy
~ n=(
v=( 0
(6.148)
Fz = 2L ~ ~ J!zvn dy
v=1 n=1 0
(6.149)
where L is the length of secondary part of motor in x-direction.

7
Computer Aided Design in Magnetics
Jan K. Sykulski
7.1 THE CAD ENVIRONMENT

Electromagnetic field computation as an aid to engineering design is an area
of very active research. The more traditional techniques used for device
analysis in practical design rely on the notion of magnetic circuit analogy, but
the computer based numerical field analysis attracts increasing interest from
existing and potential users of computational tools for a wide range of
technical applications. Many companies, in diverse branches of industry, have
already incorporated Computer Aided Design in Magnetics into their every-
day design practice. Application areas include electrical machines, MRI
magnets, electromechanical actuators, induction heating, power transmission,
robotics, non-destructive testing, electromagnetic compatibility, recording and
audio industry, transport systems, particle physics etc. The list is virtually
endless. Advances in computational techniques for design purposes, as well
as their implementations and applications, are regularly reported in learned
society journals (IEEE Transactions, lEE Proceedings), professional journals
(e.g. COMPEL [7.47; 7.71]), and specialist conferences such as COMPUMAG
[7.12; 7.13; 7.14; 7.15], ISEF [7.36; 7.71], INTERMAG, ICEM and others.
Finally, the reader is encouraged to consult other introductory and advanced
books and publications on the subject (Binns, Lawrenson and Trowbridge
[7.4], Biro and Richter [7.5], Hammond and Sykulski [7.30], Jin [7.37],
Lowther and Silvester [7.40], Silvester and Ferrari [7.64], Silvester [7.67]).
In this Chapter we shall review the more fundamental ideas and address
some of the more advanced related issues, such as optimization and error
estimation.
First, a designer has to build a computational model of his physical
problem. This, of course, is a vital step, often underestimated, which may
decide whether the whole process is a success or a failure. Clearly, the model
must be adequate for the results to be meaningful. This emphasizes the
significance of human input to the design and importance of an experiment
and/or alternative models for solving the same problem, so that comparisons
can be made. Once the model has been formulated the CAD system will
facilitate finding the solution. A typical CAD system [7.30] is illustrated in
Fig. 7.1.
There are three distinctive, although interlinked, stages in the magnetic
CAD process:
286 Computer Aided Design in Magnetics
~~m d~gn system I------~ Model

f' · ·. · ·. · ·····6Ao·l
Mesh
Solver
Post-processor
Back to design system . - - - - - { Solution
Fig. 7.1 A typical CAD system for magnetics
• Pre-processor: for defining geometry, input of material data and defi-

nition of sources and boundary conditions (induding
possible symmetries),
• Solver: for performing computation, e.g. solution of the
system of algebraic equations in the finite-element or
finite-difference formulation, or calculation of series/
parallel connections of components in the 'tubes and
slices' formulation (see Sections 1.10 and 1.11),
• Post-processor: for graphical display of field distributions, calculation
of field components, and integrating along lines, over
areas or inside volumes, to find integral parameters
such as forces, stored energy, impedances etc.
In most contemporary CAD systems both pre- and post-processing are inter-
active, thus taking full advantage of graphics capabilities of modern com-
puters. A combination of keyboard/mouse input is usually available (Le.
command driven versus menu driven), some systems offer direct transfer from
drawing packages such as AutoCad. Results files are often transferred to
mechanical stress or thermal analysis programs for further processing - such
programs may be built-in or external. The rapid developments in graphics user
interfaces (GUI) make the programs increasingly user-friendly and easier to
use. The solution process, on the other hand, is very often 'hidden' from the
user, Le. run in the background, performed using batch processing, or sent,
via a network, to a mainframe computer. (In the 'Tubes and Slices' method
and program - see references [7.31], [7.72] or [7.76] - however, owing to
The CAD Environment 287
Pre - processor
Adaptive
solver Equation solver
Solution No
OK?
Yes
.................................................................................................................................................................... ;
Post - processor
Fig. 7.2 Adaptive meshing
efficient formulation of the method, the solver is effectively amalgamated with

the rest of the graphics interface, because solution times are extremely short.)
All processors share a common database.
The engineering software used in such CAD systems is necessarily highly
specialized, and its development is influenced, on the one hand, by advances
in computer hardware and software (e.g. operating systems, window environ-
ments), and, on the other hand, by actual and perceived needs of potential
users. The life-span of any software package is usually limited. A typical large-
scale finite-element program contains hundreds of thousands of lines of code
and is the result of many man-years of development. Its total life-cycle is
estimated at about 10-12 years [7.40], which is quite a long time in a high-
technology world.
Most commercially available software is based on a finite-element formula-
tion (see the survey in Section 7.10 later in this Chapter). These programs
offer various levels of automation, but of particular interest and importance
is how they handle the mesh generation. We shall discuss various techniques
in the next Section, but it would be appropriate to mention here that in an
ideal design system the discretization should in fact form part of the solution
stage, rather than pre-processing, as shown in Fig. 7.2. This is known as
adaptive meshing, and some existing programs already offer this facility,
although perhaps it may not be yet fully integrated into the design process.
Adaptive meshing, although simple as a general concept, is rather difficult to
implement in practice, especially for time-varying fields, and is an area of
active research and development. Adaptive meshing is directly associated
with error analysis. Both topics are discussed later in this Chapter (Section
7.6).
The difficulties hindering the wide-spread usage of electromagnetic software
are that the underlying theory is conceptually complicated, mathematically
demanding and computationally challenging. At the same time, and probably
for the same reasons, electromagnetism has never been a favourite subject
amongst students of engineering and physics, and a high proportion of
designers in various sections of industry, who would otherwise benefit from
using the available CAO systems, have insufficient electromagnetic back-
ground. More effort is invested into making the subject more 'user-friendly',
and the high capabilities of modern computers in field visualization should
make an important impact on improving the situation [7.65; 7.66]. A recent
book by Hammond and Sykulski [7.30], and associated computer program
TAS, both already mentioned in this book, make an attempt to bridge the
existing gap by providing a simple and friendly computational tool, and at the
same time emphasize the field geometry and structure, rather than its
mathematical description.
At the same time researchers are looking at ways of simplifying the interac-
tion between the software and the user. One such development has lead to the
idea of Design Environment Modules (OEMs) [7.3]. A OEM facilitates the
use of electromagnetic analysis software by providing an application-specific
shell to guide a non-specialist through the geometric design and physical
property specification of a class of device. After analysis it then provides
values of the important design parameters. The OEMs are created by experts
and contain a parametrized model of a device or group of devices with a set
of decision making routines suggesting optimal representation of materials
and boundary conditions, followed by automatic or adaptive meshing. The
OEM Flowchart
OEM Data Flies 1+----,
Existing FE Software Package
User Input
Fig. 7.3 A general DEM and its relation to existing FE package

Pre-Processing and Mesh Generation 289
post-processor offers top level commands for specific tasks such as calculation
of device parameters. The command language of the software is used through-
out. A general DEM is shown in Fig. 7.3, and more details about a particular
implementation may be found in reference [7.3J. The design modules are
currently under development for the package OPERA-2d [7.48J.
The requirements imposed by a modern design environment on electro-
magnetic software are quite tough. A potential buyer or current user of a
package expects speed of computation, reliable error estimation with guar-
anteed accuracy, practical and convenient ergonomics for data input and
processing, accurate and meaningful field visualization and post-processing,
high levels of automation, fully interactive communication and reasonable
costs. Thus sophisticated and interchangeable database structures, interactive
pre- and post-processing, automatic meshing, adaptive solvers, knowledge
based systems, are all necessary components of contemporary systems for
magnetic design.
7.2 PRE-PROCESSING AND MESH GENERATION

Most finite-element programs use three types of mesh generation: deter-
ministic, semi-automatic and inter-active. The fourth possibility arises of a
fully automatic procedure when combined with an adaptive solver (see Section
7.6). The choice of method often depends upon the type of geometric
'primitives' (Le. basic geometrical shapes) which are used to construct the
finite-element model. At the same time the designer may wish to have more
control over the element distribution, for example to take proper account of
the 'skin effect'. A 'manual' input may be easily implemented for a quadri-
lateral, where the definition of boundary subdivisions results in a pre-
determined internal meshing. Some simple examples in two dimensions are
illustrated in Fig. 7.4. Clearly, there are more possibilities and the principle
applies not only to a rectangle but to any quadrilateral, which may be achieved
through iso-parametric mapping (see Fig. 7.5).
Interactive methods allow particular 'nodes' or 'construction lines' of the
mesh to be repositioned directly on the screen of the terminal using a graphics
interface, e.g. a mouse. This method is used to complement other methods,
for ~xample to apply 'finishing touches' to a manually or automatically
generated mesh.
Semi-automatic methods usually apply to polygons of any shape, whilst the
discretization of sides is either predetermined or automated using some simple
rule (e.g. similar length of the 'edge' subdivisions). The objective is first to
find the optimal distribution of internal 'nodes', and then the best way of
connecting these points. The intuitive criterion is that all triangles should be
as near to equilateral as possible, to avoid large obtuse angles, and thus to
produce a mesh which is 'well proportioned'. An illustration of a 'bad' and
a 'good' mesh for a given set of points is given in Fig. 7.6. Several algorithms
have been developed and are well covered in specialist literature. Probably the
most popular method is based on the so-called Delaunay triangulation (7 .16J.
(a)
(b)
(c)
Fig. 7.4 Simple meshes for rectangles [7.30J: (a) regular, (b) graded in one direction,
(c) graded and non-uniform
..
Fig. 7.5 lsoparametric meshing
The criterion is that the circumcircle of a Delaunay triangle may not contain
another vertex in its interior. A well constructed mesh of triangles produced
by the Delaunay triangulation maximizes the sum of the smallest angles in
the grid and thus achieves as near an equilateral set of triangles as is possible
using the given points. We shall illustrate the procedure using a particular
scheme.
During the specification of the geometry of a problem, once a region
in the form of a polygon has been defined, and boundary subdivisions entered
(either manually or carried over from another region, so that continuity of
discretization is preserved), the initial triangulation may commence. This may
be done, for example, by selecting the first two boundary points as vertices
of the first element, and then testing the suitability of all other points to
become the third vertex, see Fig. 7.7b. Two tests are conducted:
• the lines connecting the first two vertices with the new point must not
cross any existing internal or boundary lines, and
• no other boundary point is allowed on either of the two lines (as an
example, point D in Fig. 7. 7b would be rejected).
If the point has passed the two checks, the radius squared (r 2 ) of the
circumscribing circle for the newly formed triangle is calculated and stored.
The procedure is repeated for all remaining available boundary points. A point
with the smallest value of r 2 is chosen as the third vertex.
When a new triangular element is constructed, either one or two internal
lines are created. If one new internal line is formed (and thus the other one
is an existing boundary line), it is taken as the base-line for the next element
and the process is repeated, although there is now one point less available than
previously. If two internal lines are created, one is used as the next base-line
and the other is placed in a last-in first-out stack. If no suitable point can be
found, a new base-line is taken from the stack. The process continues until
no suitable point can be found and the stack is empty. An example of the
initial triangulation is shown in Fig. 7.7a. The triangular elements constructed
are of poor quality and further points need to be added inside the region.
Fig. 7.6 Meshes constructed on a given set of points /7.30j: (a) a 'bad'mesh, (b) a 'good'mesh
(a)
18
9
8
1
fj
5
4
3
2
1
8
8 1 2 3 4 5 6 7 8 9 18
(b) boundary line boundary line
point tested
i
A o
B
Fig. 7.7 Initial triangulation: (a) final result, (b) test for a new point
The procedure of adding interior points may be based on selecting the

'worst' triangular element, in terms of its shape, and inserting a new point
somewhere inside or on the edge of this element. In order to determine the
'quality' of an element we shall follow a simple criterion suggested by
Lindholm [7.38], where the quality of a triangle is measured as the ratio of
the diameter of its inscribing circle to the radius of its circumscribing circle
d
Quality Q =-
r
Fig. 7.8 Measuring the quality of a triangle
(see Fig. 7.8). Thus Q = d/r, and 0 ~ Q ~ I. Hence to find the 'worst' triangle
we need to identify a triangle with the lowest value of the Q factor.
As an example, Fig. 7.9 shows a possible strategy, where first a triangle
with the worst Q factor is found (element 5), and then a new point P is
added at the midpoint of its longest side (Fig. 7.9a). The two elements which
share this side are deleted, and four new triangles are created, as demon-
strated in Fig. 7.9b. At this stage, the new mesh may not conform to the
Delaunay criterion (for example point C in Fig. 7.9c is inside the circum-
scribing circle).
The 'swapping algorithm' tests all triangles having a vertex at P (by con-
structing the circumscribing circles), and if a case like that shown in Fig. 7.9c
is found, i.e. if the fourth vertex of the quadrilateral falls inside the circum-
circle of the other three, the existing diagonal of this quadrilateral (APBC in
our case) is 'swapped'. The final position is shown in Fig.7.9d. Any new
triangles created by the swap are placed in a last-in first-out stack. A triangle
is taken off the stack and the last step is repeated, swapping if necessary and
adding triangles to the stack, until the stack is empty. The mesh now contains
the new point P and conforms to the Delaunay criterion.
The process of adding internal points will continue until some quality condi-
tion has been achieved, e.g. the minimum acceptable value of the Q factor
has been reached by all triangles (except, perhaps, very small triangles, Le.
some 'area restriction' condition could also be in operation). The triangulation
may also be terminated manually or by setting an upper limit to the number
of triangles in the mesh. At this stage, a good mesh will have been created
that conforms to the Delaunay criterion. Further refinement is possible,
for example by applying Laplacian smoothing (node relaxation) [7.25]. As
shown in Fig. 7.10, internal nodes are 'relaxed', i.e. moved to the centroid of
the polygon formed by its connected neighbours. This has the effect of
maximizing the minimum angles of the elements forming the polygon.
(d)
Fig. 7.9 Addition of interior nodes in the FE mesh
Fig. 7.10 Laplacian smoothing
The scheme described above is based on a procedure suggested by Sloan

[7.70). It leads to a triangulation of the type shown in Fig. 7.l1a. Finally, in
order to produce a denser mesh around re-entrant corners (where the field may
vary most rapidly), a simple weighting could be applied, for example by setting
a lower value of the 'area restriction' on elements close to such corners. A
possible modified mesh is shown in Fig. 7.11 b.
In three dimensions, the extension of the Delaunay procedure is possible
(a) 18 ':'l.
1
,,- ."- >0<:::
9~~ ,I'
8
/[;::~~~~
7"'-""'1I'-J/~~
6!l)!.
""~i.v
5 /' ALI<!
,~/
'" -.p '1.)/""-1-'0-"'\/ I
3E'--,
12345678918
Fig. 7.11 Examples of final meshes: (a) without weighting, (b) with weighting
using tetrahedra and inscribing and circumscribing spheres. An appropriate

algorithm may be found for example in reference [7.9].
In addition to the various techniques for meshing polygons, most com-
mercially available programs offer useful extra features for pre-processing,
such as copying, replication, rotation, mirror image, loop input, as well as
Fig. 7.12 Finite element mesh for a Linear Variable Differential Transformer f7.75]
Code Development and Limitations 297
libraries of shapes, material properties, magnetization curves etc. Moreover,

the pre-processor usually performs several checks on the generated mesh to
ensure continuity of element subdivision, completeness of material data,
matching of interface boundaries, sufficiency of applied boundary conditions,
consistency of geometrical representation, and many other. These checks are
often highly automated. As a practical illustration Fig. 7.12 shows a typical
mesh used in FE analysis of a Linear Variable Differential Transformer [7.75J.
A combination of deterministic and automatic meshing is used.
The discussion of this section focused on geometrical criteria for mesh
generation and refinement. An addition of a field criterion, Le. a combination
of geometric and field conditions, leads to adaptive meshing, discussed in
Section 7.6.
7.3 CODE DEVELOPMENT AND LIMITATIONS

The most important, although perhaps least visible to the user, part of any
CAD system is the solution processor or solver. Many different or alternative
formulations are used for solution of particular types of equations. At the
same time the processor must be equipped with all the necessary algorithms
for discretization of the governing field equations, setting up of the appro-
priate matrix of coefficients, and finally the solution of the system of algebraic
equations. This is an area of very active research and proceedings of specialist
conferences, such as COMPUMAG [7.12; 7.13; 7.14; 7.15J, should be
consulted for up-to-date information. Another significant event in recent years
has been the series of international workshops TEAM (Testing Electro-
magnetic Analysis Methods), sponsored by the COMPUMAG conferences,
which started in 1985 as a means of aiding in development and validation of
3D eddy current codes, expanded later to include other aspects of electro-
magnetic field computation. A number of bench-mark problems have been
defined and many of them solved using special or general methods. Currently
the active TEAM problems include: coil above a crack, moving coil in a
cylinder, plate over a coil, nonlinear steel channels, rectangular slot in a
steel plate, magnetic damping in torsional mode, jumping ring experiment,
waveguide loaded cavity and microwave field in a loaded cavity. More
information may be found in the newsletter of the International Compumag
Society [7.35J.
Various possible field formulations have been discussed elsewhere in this
book, so was the importance of choosing the appropriate gauge for solution
potentials. Currently, the most popular formulations used for eddy-current
calculations are as shown in Table 7.1 [7.4J. Some other possibilities are also
explored in reference [7.5J.
The historical development of codes for solution of two- and three-
dimensional electromagnetic field problems is summarized in Table 7.2 [7.77J.
Considerable research and development are still required in a number of
areas, although significant advancements have already been made as reflected
by Table 7.2. Work currently in progress includes strategies for dealing with
Table 7.1 Contemporary eddy-current field formulations [7.4)
Method Defining equation Gauge
H-If> v x V x H = -l1o(CJHICJt)
A- V-If> V x (l/l1)V x A = -o(CJAICJt + VV) V'A=O
V'o(CJAICJt + VV) = 0 V'A = -l1oV
A·-If> V x (l/l1)V x A· = -o(CJAICJt) V·oA· = 0
T·-If> V x (l/o)V x A· = -l1(CJldt)(T - VIf» T·u=O
V'l1(T-VIf»=0
In free space in all cases If> satisfies V21f> = O.
Table 7.2 Time evolution of code development [7.77)
1960 1965 1970 1975 1980 1985 1990

D2Df 020 D2Dnp D2Da D3Dna D2Dse
Statics I2D D2Dv
I3D I3Dn
020 D2Dn· D3Dt D2De D3D
Steady state a.c.
120
020 D2Dn D3Dt D3D
Transient
120
Full Maxwell D2D D3D
Motion D2Dt D3Dt D3D
Coupled problems D2Df 020 D3D
Key: 020 differential 20 D3D differential 3D

120 integral 20 I3D integral 3D
n nonlinear a anisotropy
p permanent magnet f finite difference
s scalar hysteresis v vector hysteresis
e error analysis t restricted formulation
• approximate model t uni-directional velocity
'far field', or 'open boundary', condition (e.g. infinite elements, recursive

ballooning or mapping techniques), the analysis of moving boundaries, vector
hysteresis, coupled problems, high performance computing, material handl-
ing, adaptive solving and optimization. The reader is encouraged to consult
some of the references provided for this Chapter.
7.4 THE ART OF SPARSE MATRICES

A matrix is said to be sparse if many of its coefficients are zero and there is
an advantage in exploiting the presence of these zeros. Sparse matrices arise
in a wide range of physical and mathematical problems and are a natural
product of methods dealt with in this book. Significant savings in computer
memory requirements and computational times may be achieved as a result
of using suitable sparse storage strategies.
Consider a simple regular finite-element mesh of Fig. 7.13, where nodes
are numbered sequentially in a vertical direction and faces I and 2 have
fixed potential values (Dirichlet condition). Within each quadrilateral either
The Art of Sparse Matrices 299
()-_ _" _ _~~_ _"'::~_-(J25
)--_~ ....._ _....._ _~~_ _~24
Face 1 Face 2
O-_ _....._---..;.. ._ _-4~_~23
u-_ _..... ---:l. . ._ _. . .~_ __n22
1 21
o----...;...---...........--~ ...~--~
Fig. 7.13 A regular finite-element mesh
diagonal may be used and a particular choice is demonstrated. The mesh is

regular in the topological, rather than geometrical, sense and could fit a more
complicated shape. Potentials of nodes 6 to 20 are unknown.
In a finite-element formulation, using first-order triangular elements, this
will result in a symmetric matrix for nodes 6 to 20 of the following form
1
16 2
17 3
18 4
19 5
36 28 6
46 37 29 20 7
47 38 30 21 8 (7.1)
48 39 31 22 9
49 40 23 10
41 32 11
50 42 33 24 12
51 43 34 25 13
52 44 35 26 14
53 45 27 15
where the numbers refer to the non-zero entry positions, not their values. Only
these non-zero values need to be retained, and this can be arranged easily in
the form of two smaller arrays
1
Fig. 7.14 A finite-element mesh defined on nine nodes
123456789101112131415J
and
[ 16 17 18 19 20 21 22 23 . 24 25 26 27 .
. 28 29 30 31 . 32 33 34 35]
36 37 38 39 40 41 42 43 44 45 (7.2)
[ 46 47 48 49 . 50 51 52 53 .
The initial matrix contained 15 x 15 = 225 elements, which are now stored as
15 x 2 + 10 x 3 = 60 numbers in two convenient tables. The scheme is very
efficient; however, it is suitable only for regular meshes.
Consider now a more general case of Fig. 7.14, where nodes are connected
and numbered in an arbitrary way, and first-order triangular elements are
used. The resulting matrix will be of the following form
11
21 22
32 33
42 43 44
51 52 54 55 (7.3)
65 66
74 75 76 77
84 87 88
. 96 97 98 99
The Art of Sparse Matrices 301
II 21 51 22 32 42 52 33 43 44 54 74 84 55 65 75 66 76 96 77 87 97 88 98 99
I 2 5 2 3 4 5 3 4 4 5 7 8 5 6 7 6 7 9 7 8 9 8 9 9
I I I 2 2 2 2 3 3 4 4 4 4 5 5 5 6 6 6 7 7 7 8 8 9
Fig. 7.15 Representation of the matrix of eqn (7.3) as three linear arrays
In view of symmetry, only half of this 9 x 9 matrix needs to be stored, and

there are 25 non-zero elements. A straightforward and popular storage method
for general sparse matrices places the non-zero elements in a linear array and
complements this array with two integer arrays which store the corresponding
row and column numbers (see Fig. 7.15). In our example the storage is by
columns; storage by rows is of course also possible.
The repetition of numbers in the second integer array may be exploited by
storing only the information about where new columns begin. Thus the much
more compact modified system becomes as shown in Fig. 7.16.
The second integer array now contains pointers identifying the beginning
points of columns. If the matrix is very sparse the storage savings will be
considerable.
Another group of desirable ordering strategies are band and variable-band
(also called skyline, profile, and envelope) matrices. The matrix of eqn (7.3)
may be shown in the form of Fig. 7.17.
Space is allocated for all elements of each row starting with the leftmost
non-zero entry and continuing up to the diagonal, as shown by the shaded area
in Fig. 7.17, called the matrix profile. Some zero elements are thus stored.
However, the method is quite popular, in particular in conjunction with
Fig. 7.16 A modified representation of the matrix of eqn (7.3)
12 15
:;~;: 23 24 25
!~\;;~~ 1,1;'; 45 47 48
[It! ;~~l 56 57
Fig. 7.17 Profile storage of sparse symmetric matrix of eqn (7.3)

Fig. 7.18 Profile storage using three linear arrays
triangular decomposition algorithms. One of the reasons is that, although the

lower triangular matrix will not normally have the same pattern of zeros as
the original matrix, it will nevertheless have the same profile [7.67]. Thus the
same storage and identical addressing functions can be used for both matrices.
A possible way of storing our matrix is demonstrated in Fig. 7.18. This time
the elements (including some zeros) are stored in a linear array row by row.
The two integer arrays point, respectively, to the column number of the
beginning of each row and to the diagonal element (Le. its position in the
linear array) which forms the end of each row. Strictly speaking the second
integer array is not necessary as the ending of a row is always on the main
diagonal. However, the addressing functions become much more complicated
and a system with both arrays may still be preferred.
Most practical finite-element programs exploit matrix sparsity and the
profile storage, or its variation, is probably the most popular technique. The
space efficiency of such storage depends crucially on numbering of matrix
rows and columns, and thus on numbering of the nodes in the mesh.
Moreover, the time required for triangular decomposition of a profile-stored
matrix depends even more critically on the profile size. Hence, the initial
random node numbering is usually followed by some systematic renumbering
scheme in the search for an improved sparsity pattern. This process often
starts by classifying nodes into sets called levels [7.64]. A starting node is
assigned to level L(O), and level L(i) is made up from all available nodes
connected to level L(i - 1). Each node can belong to one level only. For
example, the levels for the mesh of Fig. 7.14 are
L(O) = [1]
L(l)=[25]
(7.4)
L(2) = [3 467]
L(3) = [8 9]
The distance between nodes is thus givep by the difference between their levels,
e.g. the shortest path between node 1 and the most distant node requires
traversing three links. The total number of levels, as well as the grouping of
nodes within the levels, depends on the starting node. In our example, starting
with node 5 yields
L(O) = [5]
L(l)=[12467] (7.5)
L(2) = [3 8 9]
The ICCO Method 303
A matrix is said to have a topological diameter of d if the highest-numbered

level reached for any starting node is L(d) (in our case d = 3). Hence a possible
strategy is to examine the matrix structure to find its topological diameter
and choose a node at either end of this diameter as a starting node. Alter-
natively, a node with the fewest connections to other nodes could be selected.
Subsequently, good numberings are usually obtained when widely different
node numbers are widely separated. Such separation may be achieved by
using the node levels in an orderly sequence, as indicated for example by
eqn (7.4). Although the amount of work involved in the search for the
'best' sparsity pattern may appear considerable, it is usually amply recovered
through decreased costs of the actual solution of the equations.
The art of sparse matrices and node numbering has its own specialist litera-
ture and a number of specific techniques exist. It is very much an area of
continuing research. The reader will find more information in books [7.18],
[7.64] and [7.67].
7.5 THE ICCG METHOD

If the coefficient matrix [N], say, is symmetric, positive definite and sparse,
as finite-element or finite-difference matrices typically are, the computational
effort may be considerably reduced by application of the incomplete Cholesky
preconditioning, followed by the conjugate gradient iterative solution. This
combination is known as the ICCO [7.45; 7.4; 7.64] and has established itself
as probably the most effective method of solving large systems of algebraic
equations of the type encountered in this book.
In the standard Cholesky decomposition, the positive definite matrix [N]
can be factorized into the form
[N] = [L] [LF (7.6)
where [L] is a lower triangular matrix with positive diagonal elements. The
elements of [L] may be found by recursive application of the following
relationships
(7.7)
for every off-diagonal element in column k in each row i, and

J
~[
;-1
L;;= N;;- ~Lt (7.8)
j= I
for the diagonal element in each row, where

L ll = "jN ll (7.9)
In the conjugate gradient (CO) method, the solution of the system of
equations
[K] [x] = [b] (7.10)

is achieved by computing successive solution estimates [x] b Le. by putting
(7.11)
where the direction vectors [p] k are selected so as to make successive resid-
uals orthogonal to each other
[r]I[r]k+l =0 (7.12)
The residual vector [r] k is given by
[r]k = [K] [X]k - [b] (7.13)
and the parameter ak is chosen to make the best possible improvement at
each step. The process starts with an initial estimate [x]o and chooses the
initial search direction [p] 0 to coincide with [r] 0:
[r]o = [K] [x]o - [b] (7.14)
[p]o = [r]o (7.15)
There follows then a succession of residuals and search directions through the
recursive step [7.64]
[P]I[r]k
(7.16)
ak =- [p]I[K][P]k
[X]k+l = [X]k + ak[p]k (7.17)
[r]k+l = [r]k + adK] [P]k (7.18)
13k = _ [p]I[K] [r]k+1 (7.19)
[p]I[K] [P]k
[P]k+1 = [r]k+1 + 13dp]k (7.20)
In the conjugate gradient method the scalar parameter 13 is chosen so that the
search directions [p hand [p h+ 1 are conjugate with respect to the matrix
[K]. Le.
(7.21)
There are alternative choices for both a and 13 and they are discussed for
example in reference [7.59].
The rate of convergence of the conjugate gradient method depends strongly
on the eigenvalue spectrum of the matrix [K]. The process may be accelerated
if an appropriate preconditioning matrix [8] is chosen and the CG method
applied to a matrix [8] [K] [8]T instead of [K]. [8] must be chosen so that
[8] [K] [8]T has a smaller eigenvalue spectrum than [K] and thus more rapid
convergence may be achieved. Equations (7.19) and (7.20) now take the form
/3 - _ [p]I[K] [8] [r]k+l (7.22)
k- [p]I[K][P]k
Error Analysis and Adaptive Meshing 305
(7.23)
and once the process has converged, the solution to the original problem may
be easily recovered.
If the matrix [K] is sparse, the incomplete Cholesky decomposition
provides a very good choice for preconditioning. The lower triangular matrix
[L] is forced to have the same sparsity pattern (the same topology) as the
lower triangular part of [K], but otherwise the decomposition is carried out
in the standard manner, except for the pre-assigned zeros. The preconditioning
matrix is then chosen to be the inverse of [L]
[s] = [L]-I (7.24)
The Cholesky decomposition is no longer exact, because of the enforced
sparsity pattern, and may be written as
[K] = [L] [LjT + [E) (7.25)
where [E] is an error matrix. The product [S] [K] [S]T then becomes
[S] [K] [sjT = [I] + [L] -I [E) [L]-T (7.26)
and if [E] is small the eigenvalue structure of the preconditioned matrix may
be quite close to that of the identity matrix, and thus the eigenvalue spectrum
has been narrowed.
Although the mechanism is not absolutely clear, it has been consistently
found that the computational effort to solve the transformed system is
considerably less than for the original system. This incomplete Cholesky
preconditioning with conjugate gradients is known as ICCG and has been very
successfully implemented in many finite-element codes, particularly for both
2D and 3D static fields.
Modifications to the method are required if the matrix is not positive
definite. One such interesting adaptation of the Cholesky decomposition is
described in reference [7.39] for coupled magnetic-electric equations.
7.6 ERROR ANALYSIS AND ADAPTIVE MESHING

Numerical solutions obtained using the finite-element method, or other
methods discussed in this book, are, by their very nature, approximate. The
difference between the numerical and the exact solution, locally in an element
or subdomain, or globally in some average sense, is the actual error of the
approximation. Such an error can be calculated only if the exact solution is
known, e.g. in the form of an analytical expression, or accurate measurement.
This is a routine phase in software development, testing and comparative
assessment, where benchmark problems are used for this purpose (see for
example TEAM activities described in reference [7.35]). Generally the exact
solution will not be known and the question arises whether the numerical
solution obtained is accurate enough. The level of acceptable error must be
set by the designer and will depend on many factors, such as tolerances of
the design, accuracy of material data and characteristics, approximations

introduced to the model, sensitivity of resultant parameters etc. A perfect
numerical solution will rarely be obtained, nor should it be required, as this
would be uneconomical in view of the many other types of errors involved.
Methods of reliable error estimation are thus required and it is currently an
active area of research. The reader will find some suggestions for further
reading in the references for this section, in particular the following publica-
tions: [7.4; 7.15; 7.17; 7.20; 7.21; 7.22; 7.23; 7.25; 7.26; 7.27; 7.28; 7.52; 7.53;
7.54; 7.57; 7.58J.
The errors involved in approximate numerical solutions are of three
principal kinds: truncation or discretization errors (although strictly these are
slightly different as discussed in Chapter 2), round-off errors due to floating
point arithmetics (Le. limited precision), and approximations in the mathe-
matical model of the real problem. Only discretization errors are considered
in this Section. They are due to the incomplete satisfaction of the govern-
ing equations and boundary conditions, and are introduced by the limited
accuracy of the trial function approximation. If the process of error estimation
is completed before the solution of the problem is attempted, it is said to be
a priori. If error is estimated after the numerical solution has been obtained,
it is called a posteriori. An interesting example of the former type is a scheme
based on a neural network used to predetermine the 'ideal' mesh density
and thus to provide an initial mesh for an adaptive system [7 .19J. The latter
method is becoming more and more popular and is usually combined with
adaptive meshing.
Intuitively, the accuracy of the solution should improve if discretization
is increased or higher order elements are used. Indeed, these are the two funda-
mental strategies for adaptive meshing, Le. mesh refinement or trial function
imProvement (e.g. hierarchical elements). In either case reliable and accurate
error estimation is required to control the procedure and make sure that
refinement is performed in such parts of the mesh where it really matters, so
that the process is efficient. We shall now discuss briefly some possible
methods of estimating the error in the finite element solution.
The first possibility is to exploit the discontinuity at element interfaces of
selected field quantities. Consider two adjacent tetrahedral elements, say 1 and
2, with a common face as shown in Fig. 7.19.
We would expect the normal component of the magnetic flux density B to
be continuous across any interface; in practice, depending on the actual for-
mulation, vector B obtained from the finite-element calculation may present
normal discontinuity. Similarly, the magnetic field strength H should have
its tangential component continuous, except in the case of a surface current,
but in the numerical solution this continuity may not necessarily be pre-
served. Hence this type of discontinuity may be used as an error estimator.
Appropriate expressions could be
error = J I(B(1) -B(2»)'nI 2ds (7.27)

s
Fig. 7.19 Discontinuity across element interface
or
error = JIn x (B(I) -B(2»)j2ds (7.28)

s
where integration is carried over the face area. As each tetrahedra has four
faces, the overall error for the element could be taken as the maximum of the
component errors, or some sort of weighted average, or simply the sum of
the four errors.
Another possibility arises if the software has the option of calculations with
both first and second order elements. Clearly the second order approximation
will be more accurate and the difference between the two solutions could be
used as a measure of the error. A convenient parameter is the energy in each
element [7.28], hence
error = Ienergy (1st order) _ energy (2nd order) I (7.29)
This approach could be easily extended to higher order elements if the code
offers such elements.
Nodal perturbations between successive mesh refinements offer another
straightforward alternative. An error indicator for each element could be
chosen as (7.32]:
error =( If,few - fO~d I + lJ;ew - f~/d I+ lJ;ew - f1Ld I + Ifn~w - foid I) X volume
(7.30)
where the indices I, 2, 3 and 4 refer to the four nodes of the tetrahedra
and f is the solution function. Care has to be exercised if, during the process
of mesh refinement, the position of nodes changes. One possibility is to
calculate f by interpolation; this, however, leads to a computationally
expensive scheme. An alternative is to solve the problem without node
relaxation, then stabilize the mesh, calculate the error and finally refine the
mesh [7.28].
A very elegant approach to error estimation exploits the dual and comple-
mentary variational formulations [7.52]. The duality is inherent in Maxwell's
equations and was mentioned before in Section 1.8. It forms the basis of the
method of 'Tubes and Slices' described briefly in Sections 1.9 and 1.10, and
more fully in publications [7.31], [7.72] and [7.76], and a book [7.30]. Com-
plementary and dual energy methods show that error bounded solutions
are obtainable when complementary pairs of functionals are extremized. In
magnetostatics, the suitable functionals have been suggested in [7.53] as
(7.31)
and
r('P) =-~J (Hs-V'P)'~(Hs-V'P)dO -J g'(nxH)dS

51
I (7.32)
where A is the magnetic vector potential, J is current density, H s is the

reduced magnetic field intensity, 'II is the reduced scalar potential, and hand
g are specified on S2 and Sit respectively. The difference between functionals
rand e indicates the degree of discretization error present in a problem. Thus,
as recommended in reference [7.54], if two solutions are available, say
standard and complementary, then the parameter chosen for evaluating error
in each element, say P(B), may be used as an error indicator by writing
(7.33)
or
error = P (B comp - Bstan ) (7.34)
where the subscripts refer to the standard and complementary solutions.
Equation (7.33) ignores errors in field direction, which may be important,
therefore eqn (7.34) is more likely to provide better error indication. It is also
possible, and may be beneficial, to weight the error indicators by element area,
for example
error = J P (B comp - Bstan ) dO (7.35)

o
The choice of the error parameter P(B) is important. An energy density
parameter, such as 1(B' H), provides one possible choice. This, however,
may lead to relatively little mesh refinement in high-permeability regions, such
as iron, because the energy density is much lower in iron than in air. A flux
density parameter such as (B'B)II2, on the other hand, would have an
opposite effect, as it enhances the importance of the flux in the iron in the
final solution. Another approach is to weight the parameters by the element
area. The main drawback of the complementary scheme is that it requires two
finite-element solutions at each stage.
An interesting alternative derivation of complementary variational prin-
ciples is put forward in reference [7.61J. Here the constitutive relationship
B = ~H is used directly and is cast in error form. This assumes the existence
of a constitution error, A, having the following properties
A~O (7.36)
with
B= ~(H)H + Br]
A = 0 <* and everywhere in R (7.37)
[
H = v(B)B + He
where Br and He are known constants. Hence A is non-negative, and is zero,
Le. minimum, if and only if the fields Band H satisfy the constitutive
relationship everywhere in the region. Such an error form provides a
variational principle, so that we can solve
cSA =0 (7.38)
to minimize A and thus obtain the fields. A suitable expression for the
constitution error in linear media is
(7.39)
where the brackets ( > indicate integration through the region of interest. It
has been shown in reference [7.61J that for well-posed problems A splits into
complementary H- and B-system functionals; these provide complementary
variational principles that may be solved independently of each other. The
error, which is wholly attributed to the numerical constraints in an approx-
imate solution, is always defined and computable locally and globally.
There exist less expensive methods based on a single solution and utilizing
the polynomial interpolation theory. For example [7.4J, using the first order
finite elements with linear shape functions, an element average error can be
evaluated by computing the integral over the element of the difference between
the nodally interpolated fields Bs (using the first order element shape func-
tions) and the element constant fields B (obtained directly from the potential
solution). The local error thus becomes
error = J (B - Bs ) dO (7.40)
o
This technique uses the information provided by neighbouring nodes to
evaluate the magnitude of the higher order terms in the solution that have been
neglected. In another paper [7.20J, the error estimation is based on the
evaluation of the term
g = Igradlgrad vii (7.41)

where V is the computed finite-element solution. It is also suggested that the
gradient of the solution is not calculated by differentiating the first order
shape functions, but is obtained by assigning to the nodes field values calcu-
lated using the five-:links generalized finite difference method suitable for
irregular grids and linearly interpolating them on the elements. In the same
paper other single solution methods are discussed, one based on approximate
solution of a local boundary value problem with hierarchical shape functions,
and two which utilize a residual evaluation.
The objective of an adaptive mesh generation is to produce an optimum
mesh automatically, i.e. a mesh which guarantees the required level of
accuracy achieved in the minimum time. This process is illustrated in Fig. 7.20.
A simple and practical approach has been suggested in reference [7.28].
Only elements with errors greater than a threshold value are refined. If the
distribution of errors has a typical form like that shown in Fig. 7.21, then all
elements on the right of the threshold line will be refined. A good threshold
value has been suggested to be the mean value of all element error indicators
possibly weighted by a constant. The procedure leads to an efficient number
of elements being refined at each step.
To illustrate the implementation of the above algorithm, a problem of
current flowing through the conductors junction, as used in metal welding,
has been solved [7.78]. Figure 7.22 shows different levels of discretization
BEGIN
PROBLEM DEFINITION
FORM SPARCITY PATTERN
FORM SOLUTION MATRIX STABILIZATION CYCLE

Delaunay transformations
SOLVE EQUATIONS Node relaxation
ERROR ESTIMATION MESH REFINEMENT
ADAPTIVE STRATEGY
Criteria Refinement indicator
No Convergence parameter
for termination
satisfied
Fig. 7.20 Self-adaptive refinement algorithm

Number of
Tetrahedra
Threshold value
Error
Fig. 7.21 A typical distribution of element error
achieved by the adaptive refinement procedure and Fig 7.23 the resultant
distribution of equipotential lines.
A particular strategy favoured by a group of researchers from Genova in
Italy [7.17; 7.21; 7.22; 7.23] uses a linear combination of dimensionless
ratios, with suitable weighting factors, to introduce two key quantities: the
refinement indicator and the convergence parameter. A standard quadratic
norm is used, which for a generic scalar or vector quantity is defined as
(7.42)
The refinement indicator is then specified as
(7.43)
where cPj is the computed solution over the element j, VcPj is its gradient (or
curl), e; is the estimate of the error on the solution over the element i, Ve; is
the estimate of the error of the gradient of the solution over the element i,
N is the total number of elements of the domain and K is the weighting factor.
The convergence parameter is defined as
(7.44)
(b)
(c)
Fig. 7.22 Discretization of the 'electric welding' geometry /7. 78J: (a) 28 nodes, 36
tetrahedra; (b) JJJ nodes, 420 tetrahedra; (c) 602 nodes; 3025 tetrahedra.
and the adaptive process is stopped when c falls below the user defined value.
For K = 1, Le. when only the estimate of the error on the gradient (or curl)
is used, this reduces to
N
2: II Veil1
i=1
2
c= N
(7.45)
2: IIVcf>jI12
j=1
The refinement criterion is that the element is subdivided if

T1i > ~T1max (7.46)
where T1max is the maximum value of the refinement indicator TJi for all
elements, and ~ is the threshold value, 0 < ~ < 1. In a modified version [7.23J,
Optimization 313
Fig. 7.23 Distribution of equipotential lines f7. 78J
two threshold values are introduced, 0 < ~l < ~2 < 1, and the ith element is
subdivided once if
~I 1'\max < 1'\; < ~2 1'\max (7.47)

and twice if
1'\; > ~2 1'\max (7.48)

The above procedure may be used with various error estimators mentioned
before.
Methods relying on adding (subdividing) elements to the existing mesh are
sometimes called the h-type; an alternative strategy involving increase of the
order of element representation is known as the p-type. Combined algorithms,
called the h-p type, are also emerging. The h-p adaption may be either
decoupled (h- followed by p-adaption) or fully integrated. Some recent
advances in all three methods are reported in reference [7.15].
7.7 OPTIMIZATION
Optimization of geometric boundaries and placement of conductors to achieve
specific design goals is paramount in computational magnetics. The under-
lying theory and mathematical formulations are well developed and docu-
mented in specialist books, see for example reference [7.24]. Increasingly,
investigators are exploiting optimization theory and apply it to magnetic
designs in order to achieve automatically specified objectives. Many proven
methods and techniques exist, such as direct-search methods, evolution
algorithms, steepest-descent and conjugate gradient methods, Newton and
quasi-Newton schemes, stochastic methods, simulated annealing techniques,
and many other, not to mention variations and modifications. A selection of
solved problems in electromagnetics may be found in references [7.62], [7.63],
[7.68] and [7.69], or more generally in the proceedings of COMPUMAG
conferences [7.12-7.15]. Simple optimization subroutines are often offered as
part of a development software, such as MATLAB [7.44]; the main task is
then to link such procedures with the magnetic field analysis software.
As an example, consider a typical solenoid actuator shown in Fig. 7.24.
A suite of programs called EAMON (Electromechanical Actuator Model
core tube frame coil
axis of symmetry annature pole face
Fig. 7.24 Radial cross-section of a solenoid actuator
Optimization [7.11]) has been developed to interface with a general purpose

finite-element package OPERA-2d [7.48]. The optimizing algorithm employs
a mixture of conjugate gradient and descent methods with Marquardt modi-
fications [7.68] for minimization in a least squares sense of a set of objective
functions.
FEM package OPERA-2d

Required force characteristics f,
and design variables Z
Parameterized model
database
PRE
Creation 01 model
,.
database files at different
armature displacements
,.
Request vector f
(force for various SOLVE Equation solver
BrmBtUf9 disp/Bcsments) I I
Calculation of co-energies
POST at various armature
displacements
NO
G
minimum
?
Sub-program to compute forcss

(using virtuBi work principle)
Fig. 7.25 Flowchart of the optimization program EAMON [7.ll]

Optimization 315
Force (N)
70,---------------,
40
,,
30
x required \
\
- optimized
20
._. initial
10
O'-------'--------'-----l..-_ _---J
o 1 234
Displacements Cmm)
Fig. 7.26 Optimization of an electromechanical actuator
The field modelling of the actuator has been conducted using a customized
shell developed for OPERA-2d [7.10]. The shell is designed for Sun Open-
window (CusOP) and automates the entire process of magnetic design. Force
calculations are based on the Virtual Work Principle (VWP) and calculation
of co-energy, although the Maxwell Stress Tensor (MST) method has also been
found helpful. The flowchart of the optimization program EAMON is shown
in Fig. 7.25.
The optimizer operates on the objective functions which are selected as the
difference between the required force curve, I" and the calculated force, I,
using the VWP method, for a range of displacements, i.e.
r=I-lr (7.49)
The function G is then defined as

N
G =~ (wrY (7.50)
;=1
where N is the number of armature displacements and w the weighting. The

task is to find the variables set Zopt for which OCZ) is minimum. Because of
the nonlinearity of OCZ) an iterative approach is employed. As an illustration,
Fig. 7.26 shows a force/displacement characteristics of an optimized actuator,
where the optimization has been applied to the armature length.
7.8 POST-PROCESSING AND FIELD VISUALIZATION

Post-processing is the art of extracting useful and meaningful engineering
information from the numerical results of a problem solution. The solution
is usually provided in the form of a set of values of a potential (scalar or
vector) calculated at a large number of nodes of a mesh, and thus it needs
to be processed before it can be viewed or related in any way to design para-
meters of the device. Postprocessing techniques typically require that the
problem is set up in a particular way. Thus pre- and post-processing are in
a way inseparable, they usually share a common database and in contem-
porary commercial software they are often combined to form one large sub-
program (see for example the Pre- and Post-Processor of OPERA-2d [7.48]
or PC- OPERA [7.51]). For an engineer the post-processor is an invaluable
interface for integrating the magnetic field computation into a design process.
One of the more familiar ways of displaying the solution to a magnetic field
problem, in particular in two dimensions, is through contour plots. If vector
potential A is used as. a solution function, Le. A, in 2D plane or A e in 2D
axisymmetric systems, then contours of constant A, or rAe, respectively,
become tubes of magnetic flux, and if they are equidistant in value then the
same amount of flux is always contained between neighbouring lines. This
type of field display is particularly helpful to designers of magnetic devices.
As an example, consider a simple case of a magnetic hollow cylinder inserted
into an applied uniform transverse field. This problem has an analytical solu-
tion, derived for example in reference [7.30]. A typical magnetic field distribu-
tion is shown in Fig. 7.27.
This type of a magnetic tube is often used to describe magnetic screening
effects, and indeed we can see how the field has been effectively excluded from
the interior of the tube. If magnetic nonlinearity comes into consideration, or
if 'slits' or 'holes' are made in the tube, analytical solutions are no longer
=----=-========-=---=
---=============-=---
Fig. 7.27 Screening effect of an unsaturated iron tube
Post-Processing and Field Visualization 317
(a) (c)
(b) (d)
Fig. 7.28 Field distribution in the magnetic tube: (a) unsaturated tube, (b) magnetically
saturated tube, (c) tube with a slit at 0°, (d) tube with a slit at 90°
available, but numerical analysis presents no difficulties. Figure 7.28 demon-

strates some particular interesting cases; owing to symmetry only one quadrant
of the solution has been shown in each case. These displays are useful as they
demonstrate, in visual form, both qualitatively and quantitatively, important
design principles for this type of screening. In this case they point to the
dangers of impairing the screening performance by saturating the tube wall
(case (b» or by putting the slits in the wrong position (case (d». Case (c), on
the other hand, shows that if a more appropriate position of a slit is chosen
it will hardly affect the screening. By counting the contour lines entering the
interior of the tube it is even possible to estimate the degree of the worsened
performance of the screen, although for this purpose a more accurate result
may be obtained by other methods, for example through direct interrogation
of field levels in different regions. In practice different ways of displaying
results are often combined and cross-referenced.
A second example is for the Linear Variable Differential Transformer
described previously in Fig. 7.12 and based on solutions presented in reference
[7.75]. Field plots for two positions of the inner core are shown in Fig. 7.29.
Contour plots are frequently complemented by displays, usually in the
form of coloured zones, showing distribution within regions of a specified
(a)
(b)
Fig. 7.29 Magnetic field plots for a Linear Variable Differential Transformer: (a)
displacement = 0, (b) displacement = 0.5 mm
component. Values of interest are typically magnetic flux density, induced

current density, permeability and error distribution (if error is calculated).
Unfortunately, such coloured 'maps' do not reproduce well in black and white,
and are not induded here for illustration.
Another useful and informative way of displaying results is to show a
distribution of a field component along specified lines (straight or curved).
The component is calculated at a number of points along the line and shown
as a piece-wise linear function, or first interpolated or smoothed and then
displayed.
In three dimensions, field visualization is much more complicated. Two-
dimensional techniques are used to some extent, i.e. fields are displayed on
identified or specified surfaces or planes. Again, colour is usually very helpful.
Line graphs are also used. Fully three-dimensional displays are much harder
to achieve. Researchers are trying various methods and one of such interesting
attempts is described in reference [7.80], where a generic 'cube' is defined on
part of a solution and then various field displays are applied to this cube.
These include coloured contour maps on a series of planes defined on the cube,
splitting the cube into sub-cubes with colours determined by field values at
the particular location, and transparent displays on planes parallel to cube
walls. Probably the most popular method of showing the fully three-
dimensional field patterns is based on the idea of a three-dimensional arrow
shown in perspective (see Fig. 7.30). The size of the arrow is proportional to
the magnitude of the field quantity. Thus both the value and the direction of
a vector quantity are indicated on the display.
Examples of images of a three-dimensional arrow are shown in Fig. 7.31.
Cases (a) and (b) refer to some general 'positive' and 'negative' directions,
Post-Processing and Field Visualization 319
Fig. 7.30 A three-dimensional arrow representing a vector quantity
a) c)
b) d)
Fig. 7.31 Examples of positive and negative images of the three-dimensional arrow:
(a) positive image, (b) negative image, (c) positive degenerate image, (d) negative degenerate
image
respectively, whereas cases (c) and (d) show particular positions of the arrow,
when it is perpendicular to the plane of display and pointing outwards or
inwards.
A popular variation of the three-dimensional arrow uses a circular rather
than rectangular base for the arrowhead. Figure 7.32 shows how such arrows
may be used to describe the eddy-current distribution in a conducting cube
immersed in a homogeneous harmonic magnetic field [7.78]. This particular
solution has been obtained using an adaptive procedure applied to the skin-
effect problem, which led to a mesh consisting of 3100 tetrahedra and 1008
nodes.
Fig. 7.32 Eddy-current distribution in a conducting cube
Modern post-processors offer much more than post-viewing and are capable
of manipulating the field solutions. Hence global parameters such as stored
energy (and thus inductance or capacitance) or ohmic power loss (resistance)
may be calculated via volume integration of appropriate field components.
Surface integration, on the other hand, is required for finding stresses, and
thus for calculation of forces and torques. In order to improve the accuracy
of quantities whose computation relies on differentiation of the solution
potential, smoothing techniques may be applied to refine the quality of the
output data. Finally, specific values or their distributions may need to be
prepared in a particular way if transfer of output data is required to another
program or output from magnetic design is going to be used as input to
mechanical stress or thermal analysis programs. For example, forces of
electromagnetic origin will constitute sources of a mechanical stress field, or
power loss in conducting regions will give rise to a thermal field. These are
not yet fully coupled fields as there is no feed-back loop; nevertheless in many
practical problems a single sequential solution is quite sufficient as the change
of magnetic or electric properties due to mechanical stress or change in
temperature may be negligible. Some commercial programs (e.g. OPERA-2d
[7.48]) have such stress or thermal analysis solvers built-in as extensions to
the electromagnetic solving procedures.
Error analysis and adaptive meshing add to the variety of tasks which may
be required during the post-processing stage. Errors have to be calculated first,
using perhaps one of the techniques discussed in Section 7.6, and may be
interrogated using available display options. If a mesh refinement procedure
is then activated and information and decisions fed back to the pre-processor,
then an adaptive meshing loop has been completed.
The Impact of Parallel Computing 321
Finally, optimization techniques come into consideration. The post-

processor then provides a feed-back link where design parameters are retrieved
from the magnetic solution and passed back to the optimization algorithm.
The pre- and post-processor become even more fully integrated, as demon-
strated before in Fig. 7.3 in relation to the Design Environment Module or
Fig. 7.25 for the optimization program EAMON.
7.9 THE IMPACT OF PARALLEL COMPUTING

Until the early 1970s most computers were strictly serial, i.e. one arithmetic
operation was completed before the next commenced. The execution time of
a program implementing a numerical algorithm could be approximated by the
formula
time = (average time per operation) x (number of operations) (7.51)
In recent years, however, many new advanced computer architectures have
been developed which employ one or more degrees of parallelism. For exam-
ple, arithmetic operations may be segmented into several distinct phases and
functional units designed so that different operands can be in different
segments of the same operation at the same time. This technique, called
pipelining, is employed by machines commonly called vector processors, since
pipelining is particularly useful when performing calculations with vectors.
The model for vector computation can be written as [7.18]
time = (n + s)c seconds (7.52)
where n is the length of a vector, s the number of segments, and c the clock
cycle time in seconds. The time sc before any result is produced is usually
called the start-up time, and s is sometimes termed the nl/2 value. This is
because the rate of computation attains half of its peak performance of 1/c
(sometimes denoted roo) at a vector length of s. With the nJI2 and roo notation,
formula (7.52) becomes
time = (n + nJ/2)/roo (7.53)
The present generation of parallel computers appear to be based on three
main architectural characteristics [7.8; 7.77]: memory structure, grain size and
control. Memory can either be distributed, with each processor controlling its
own memory, or it can be shared among several processors. Grain size refers
to the computational power of each node (each processor and its associated
memory). Control can either be SIMD (Single Instruction, Multiple Data) in
which the processors execute identical instructions in lock-step, or MIMD
(Multiple Instruction, Multiple Data) in which the processors run different
codes asynchronously. Examples are:
• distributed memory, small grain size, SIMD machines like the Digital
Array Processors (DAPs),
• shared memory, small grain size SIMD machines like the family of
vector processors,
• shared memory, large grained MIMD machines like the CRAY-2,

• distributed memory, large grained MIMD machines like the INMOS
transputer.
The implementation of finite-element codes on parallel machines is a natural

development that has been accelerated by availability of a variety of parallel
architectures. Much of the effort goes into parallelizing of the solution of the
system of equations, although element and matrix assembly have also been
implemented. An interesting comparison is presented in reference [7.33). First,
results are reported obtained using the Massively Parallel Processor (MMP)
which is a very large SIMD processor. The speedup in solution time was
obtained by reducing the number of sequential operations and exploiting the
intrinsic parallelism of the Gauss elimination and the Gauss-Jordan algo-
rithm. While the number of operations on sequential computers is of the order
of n3 , where n is the number of equations, the number of operations on the
MPP is of the order of n3/N 2 , where N 2 is the number of processors. For the
MPP, this translates to an ideal speedup of 16384. In practice, the figure is
much lower and depends on the communication required, as well as on the
size of the problem. Improvements of up to three orders of magnitude have
nevertheless been reported.
In the second comparison, solutions were obtained using a MIMD com-
puter, where each computer works independently, and thus efficiency of
processors is very important. For this experiment the following definitions
were used
d CPU (one processor)
(7.54)
spee up = CPU (N processors)
while efficiency was defined by the ratio

. .
effiCiency = speedup
N x 100010 (7.55)
The results have shown that the speedup is almost linear with the number of
processors. At the same time efficiencies ranging from 70% to 98% were
achieved, depending on the problem. For larger size problems, an almost
constant efficiency ranging between 80% and 90% with up to nine processors
has been obtained. These results are very encouraging and clearly demonstrate
possible benefits of parallel computation.
Parallel computing is a rapidly developing branch of computer science and
its potential impact on CAD in electromagnetics cannot be overestimated. As
a demonstration of a very recent initiative, PAFEC - one of the leading
developers of general purpose finite element analysis software packages - and
the Parallel Applications Centre at Southampton [7.50), are collaborating
to produce a parallel version of the PAFEC-FE finite element package.
A generic, scalable prototype Parallel PAFEC-FE made its first public
appearance at Supercomputing'93 conference in November 1993. The proto-
A Short Survey of CAD Systems for Magnetics 323
type allows parallel computation to improve performance for both static and
dynamic structural analyses. The parallelized code is fully integrated into the
original sequential package, with parallel efficiency over 90070. Substantial
improvements in turn-around time for complete analyses have been obtained
for 16 processor runs for models with as few as 15 000 degrees of freedom.
Larger models and greater number of processors will no doubt yield even
better performance gains.
7.10 A SHORT SURVEY OF CAD SYSTEMS FOR MAGNETICS

This section is not intended to provide a full catalogue of all available software
in electromagnetics. Nevertheless, it seems worthwhile to include some very
brief information about programs which perhaps have made the greatest
impact on the progress in the field of computational magnetics. There now
exist quite a few commercially available systems offering integrated tools for
CAD in magnetics.
A typical commercial package will have most of the following components:
• pre- and post-processor: fully interactive with colour graphics, windows

environment, sophisticated post-viewing facilities, comprehensive range
of supported output devices, automatic and adaptive meshing;
• statics: magnetostatics and electrostatics analysis with nonlinear
materials, including permanent magnets, special versions for laminated
materials using a 'packing factor';
• steady-state eddy currents: steady-state a.c. eddy-current analysis,
including complex permeabilities, approximate nonlinear solutions
(fundamental harmonic field), background d.c. fields, external circuits
for voltage-driven problems;
• transient eddy currents: full transient eddy-current analysis, with non-
linear materials, multiple drives and background d.c. fields;
• motional eddy currents: uniform motion induced eddy-current analysis
(with constant or varying topology);
• stress and thermal: mechanical stress using forces, or thermal analysis
using ohmic heating, calculated from electromagnetic solutions;
• 2D, 2D axisymmetric and 3D formulations.
• implementations on a variety of hardware platforms, including 386/
486/Pentium DOS machines, HP, Sun SPARC, Apollo, DEC and IBM
workstations, and Cray supercomputers.
Two particular commercial systems will be described here in some detail,

but several other will also be mentioned. Finally, examples will be given of
in-house systems, in particular those used in academic and research institutes.
Such systems are often very advanced and use state of the art formulations
and new developments. However, they usually lack the robustness and
reliability of commercial systems and the graphic interfaces are usually much
less sophisticated.
Fig. 7.33 User interface in MagNet 5
One of the oldest electromagnetic analysis software systems on the market,

available commercially since 1978 from Infolytica Corporation in Canada and
Infolytica Ltd in the UK, is a suite of programs called MagNet. The latest
release is MagNet 5 [7.41; 7.42; 7.43]. The package employs a modular
approach and a common operating environment on all hardware platforms.
Hence, the user sees a similar interface, and has all software facilities avail-
able, on hardware ranging from a PC386 to a SuperComputer. Two special
modes of operation are available, known as Fast-Track and Tool-Box.
Fast-Track allows new users to gain quick access to the package and offers
experienced users a quick start on more advanced problems. It is entirely
windows-based (see Fig. 7.33), and input is via mouse-button selection,
together with data entry in dialogue boxes, where required. An object-oriented
approach to model building is employed, whereby both simple and compound
closed regions with associated material properties, known as devices, may be
stored for future use. The ability to create new devices from assemblies of old
ones, together with a scaling facility, gives a rapid and powerful model-
building capability. An electric-circuit facility allows devices to be inter-
connected, or to be connected to external sources and components. Fast-Track
is applicable to 20 or axisymmetric problems, employing vector or scalar
potentials, and solves static and linear time-harmonic problems. Variable
element order is available, up to fourth. To further simplify the process
adaptive meshing is an option, whereby repeated solutions are performed,
Po~l~.'
MOlO _F
J,III••_. e--r-_ _.....

_ _..!PN:! .!:!. N_ _ _ 1$_1A_"_:_-_"_·_._..
1
_~_-_45 ~[
o.sAJ I:::J
HdlI
F_
FT.......
Resu.lt:.
1751 48l3E-O
Post:. 5 1801
Fig. 7.34 Postprocessing in MagNet 5
with mesh refinement being undertaken in regions where fields are seen to
be varying most. The Fast-Track Post Processor detects the problem type, and
selects the appropriate menus and display characteristics. It then gives access
to all standard line, arrow, and density plots. Point values of all variables are
available, as well as displays of quantities along contour lines, or enlarged
displays over local areas (zones). Global quantities, such as energy and its
related quantities, such as force and torque, are also available at the click of
a button.
Tool-Box makes available script facilities to run multiple problems in batch
mode. A full set of macros is also provided, together with the facility for the
user to develop new ones for specific requirements. All the modules are
available directly in Tool-Box, covering mesh generating, problem defining,
post processing (Fig. 7.34), as well as file managing. Four 3D solvers are also
available in the Tool-Box. PM3D covers 3D magneto- and electrostatics, and
TH3D is the corresponding linear time harmonic solver. TR3D is a 3D electro-
magnetic transient solver allowing voltage or current input of arbitrary wave-
shape, and interconnection of resistors or inductors. Both linear anisotropic,
and nonlinear isotropic materials may be modelled. HF3D is a high frequency
3D solver, designed to calculate the scattering parameters of passive micro-
wave components of arbitrary shape. Lossy dielectrics may be included, and
Fig. 7.35 User interface in OPERA-3d
boundary conditions available are perfect electric and magnetic conductors.

Typical quantities calculable are energy density and the Poynting Vector.
MagNet 5 employs the robust solver technology developed over the years
by Infolytica. In particular, the 3D solvers largely employ accurate and very
fast scalar-potential edge-elements, as well as hierarchical techniques which
vary element order, giving the finest detail in the most critical regions.
Increasingly, MagNet 5 appears as an invisible inner core of a much larger
package, for which the input comes directly from the designer's standard data
sheet. This is referred to as the shell mode of operation, and this variety of
modes of usage is believed to be unique to the MagNet system.
Another famous name in software for electromagnetics worth mentioning
in this Section, is associated with the company Vector Fields Ltd, Oxford, UK,
and Vector Fields Inc, Aurora, USA, and a suite of programs called OPERA
[7.48; 7.51; 7.79]. This includes several programs: OPERA-2d, and its version
for personal computers called PC-OPERA, with integrated pre- and post-
processing, for solution of two-dimensional static, eddy-current and transient
fields, including axisymmetric problems; OPERA-3d pre- and post-processor;
TOSCA for solution of static three-dimensional fields; ELEKTRA for three-
dimensional eddy-current problems. Vector Fields is well established in the
international market and the software is highly regarded for its extensive func-
tionality, versatility and good customer support. The software is command
and/or menu driven and a typical user interface is shown in Fig. 7.35.
The most recent release of the OPERA software [7.79] contains many new
features, mainly improvements and enhancements to user interface and com-
munications. For example, a new interface to the Autocad CAD package
enables geometric data to be transferred to the OPERA-2d pre-processor.
Both 2D and 3D solvers can now cope with high speed motion using a
numerical technique known as 'upwinding'. A new analysis program for
modelling beams of charged particles, including the effects produced by space
charge, has also been introduced. Simplified construction of models has been
achieved in a 2D pre-processor through the introduction of a 'background'
region (for example filling in the 'empty' space with air) and automatic region
matching. Finally, an adaptive solver has been introduced for 2D static
fields.
This new adaptive process uses a 'three step' refinement procedure. First,
the adaption is implemented, where the refinement of an element is based on
the local error criterion. An improvement process is then applied, where two
new criteria (based on area and length) are introduced to avoid· deformed
elements. Finally a regularization process is activated, where for each internal
node a new optimal position is found at the centre of gravity of the polygon
formed by the neighbouring nodes. A new solution is found for the refined
mesh, and the process is repeated until the error in each element is less than
that specified by the user, or the number of element exceeds a maximum value,
or the maximum number of iterations have been performed. In addition, a
limit is placed on the smallest element allowed. To reduce the number of
refinement iterations the value of the potential for each new node is inter-
polated from the surrounding nodal values. Thus the process always starts
from a previous non-zero solution. As an example, Fig. 7.36a shows the initial
mesh for a motor problem, and Fig. 7.36b demonstrates the final mesh (and
flux lines) after the adaptive refinement process has been completed.
The following list of other popular commercial systems is not exhaustive,
but is provided here for reference.
• ANSYS [7.1] - developed and maintained by Swanson Analysis Systems,

Inc. (SASI) of Houston, Pennsylvania. One of the oldest commercial
finite-element systems, ANSYS began in 1970 as a structural and thermal
analysis package. Magnetic solvers were first introduced in 1982. Of
particular interest are the coupled magneto-structural-thermal field
analysis options and solid modelling capabilities of the software. A PC-
based subset of the ANSYS program, the ANSYS-PC/MAGNETIC,
was designed specifically for electromagnetic analysis and has been
available since 1991.
• COSMOS/M - by the Structural Research and Analysis Corporation,
specialists in desktop finite-element analysis software. The package
includes electro-thermal and magneto-structural couplings.
• Maxwell - by Ansoft Corporation, Pittsburgh, USA. The Maxwell 2D
Simulator can run under Microsoft Windows 3.1 graphics interface.
(a)
'-- ....J
\EOPERA·2d
... _ .....
~_1.
(b)
~OPERA-2d
'-- -' 11.-.,." ....."*'-1"
Fig. 7.36 Adaptive meshing using OPERA-2d [7. 79J: (a) initial mesh for a motor problem,
(b) final mesh after refinement process
• MSCIEMAS - by the MacNeal-Schwendler Corporation, who also

developed MSC/NASTRAN. This 2D and 3D solver is particularly
useful for solving problems in electromagnetic imaging and com-
patibility, radio frequency shielding, microwave heating, magnetic
signature, antennas and other radiating structures and is available on
most popular workstations and larger computers.
• BEASY (7.2] - The Wessex Institute of Technology, Southampton, UK,

is recognized as the world's leader for boundary element technology, and
its software development arm, the BEASY Computational Mechanics
Company, offers very advanced software for a variety of applications.
The range of analysis capabilities includes stress, thermal, mechanical
fracture, acoustics, potential flow and electro-magnetic analysis. One of
the advantages of using boundary elements is simplified modelling as
only surfaces have to be defined. Advances in the boundary element
method are regularly reported in professional conferences (see for
example references [7.65] and [7.66]) and are the subject of specialist
books (e.g. [7.6; 7.7]).
• Integrated Engineering Software [7.34] - with a suite of programs
ELECTRO (2D), OERSTED (2D), MAGNETO (2D), COULOMB (3D)
and AMPERES (3D), all based on a boundary element method. Each
program is fully integrated around a convenient user-interface and
includes geometric modellers and appropriate field solvers. Geometry
files may be exchanged with other CAD packages.
• MEGA - by Bath University Applied Electromagnetics Centre, UK.
MEGA is a very comprehensive suite of 2D and 3D electromagnetic
programs, with many unique formulations, especially for transient 3D
eddy currents and moving media.
• SLIM - an integrated collection of software modules developed by GEC
Alsthom Engineering Research Centre, Stafford, UK,
• plus a number of other packages.
From the many in-house systems developed in academic and research

institutions, two will be mentioned here.
• SONMAP (7.29] - This is an advanced suite of programs, written at the

Technical University of Szczecin, Poland, incorporating many state-of-
the-art developments in finite-element techniques, such as adaptive
meshing, infinite elements, surface impedance boundary condition etc.
The package is available commercially and may be supplied in a source
code version.
• ARISTOTLE-EUKLID (7.78] - ARISTOTLE is a research package
developed at the Aristotelian University of Thessaloniki. Simplex ele-
ments are employed with node and edge shape function expansions for
solution of static and time varying fields at both low and high frequency.
EUKLID implements a three-dimensional adaptive refinement on tetra-
hedral meshes. Examples obtained from these programs have been
shown in this Chapter as Figs. 7.22, 7.23 and 7.32.
There is a continuously growing market for electromagnetic software as more

and more design offices in many sections of industry invest in Computer Aided
Design of electromechanical devices. Thus further improvements to the existing
finite-element and other codes, as well as new CAD systems, are to be expected.
8
Experimental Methods
Kazimierz Zakrzewski
8.1 INTRODUCTION
Experiments are the prime source of information about physical processes
and are conducted to assess material properties, identify equivalent para-
meters or formulate empirical relationships. Experimental investigations are
indispensable as a means of verifying various computational methods and
models. Operational characteristics of devices are usually established through
tests and measurements and they provide an ultimate answer to whether the
design objectives have been fulfilled.
In this book we have discussed various approximation methods used to
obtain field distributions via the solution of differential or integral equations.
Such solutions may also be obtained by mathematical or physical analogy
using analogue computers or special analysers.
The final verification of the design is performed by investigating prototypes.
These are usually objects with a high degree of complexity both in terms of
geometrical configuration and in view of the electromagnetic, mechanical or
thermal processes.
8.2 PHYSICAL MODELLING

Experimental investigations on real engineering objects, especially high power
electrical devices, may be very expensive. Moreover, from the technical
point of view, it is not always possible to prepare a versatile programme of
investigations which is applicable to all cases. Therefore, in addition to
mathematical modelling, physical modelling continues to play an important
role. Measurements are taken on a model of reduced dimensions but similar
physical properties and configuration. The relevant theory enables the results
obtained on a small model to be transposed to the full-scale original device,
without the need to experiment on a large prototype [8.10; 8.22J.
8.2.1. Scaling criteria

A set of dimensionless numbers resulting from the application of fundamental
equations of electromagnetics to both the original system and its model,
provide the scaling criteria for physical modelling [8.10; 8.22J.
Maxwell's equations applied to the original system, taking into account the
velocity term, yield
Physical Modelling 331
aD
curlH = oE - at + o(v x B) + J; + vdiuD + curl(D x v)
curlE = - -
aB - curl(B X v) (8.1)
at
B=IJ.H, D =EE
For the reduced-size model

mH mtmE aD
-
m,
curlH = momEoE + - - -;- + mOmVmlJ.mHo(v X
m ut
B) +
r
mjiJ; + mvmtmEv diu D

mE mlJ.mH aB mlJ.mHmv ( )
- curl E
m,
=- - m
- -at - m,
curl B X v (8.2)
r
The quotient of a given physical quantity in the model to its value in

the full-scale original is called the scaling factor and is designated by the
letter em' with an appropriate index. For example m, denotes a scaling
factor of a linear dimension. If we introduce two further self-explanatory
relationships
and
we can clearly see that equivalence of Maxwell equations for the model and
the original necessitates the fulfilment of the following conditions
momEm/m H =1 (8.3)
mtmEmfm/mH = I (8.4)
momlJ.mfm;= I (8.5)
=I
mj;m/mH (8.6)
mlJ.mHmfm/mE = 1 (8.7)
Rearranging terms leads to a set of scaling criteria
1t 1 = mlJ.momfm; = I (8.8)
1t2 = mtmlJ.m}m; = 1 (8.9)

1t3 = mj;m/mH = 1 (8.6a)
It will be noticed that in practice it may be difficult or virtually impossible
to meet criteria (8.8) and (8.9) simultaneously. In particular, assuming the
decrease of dimensions coupled with increase of frequency, and in addition
keeping mt = mlJ. = I, we obtain a condition m o = mf' which in general is
an impossible requirement, because a material of the postulated electric
332 Experimental Methods
conductivity is unlikely to be available. Hence, for eddy-current problems,

we often consider the criterion (8.8) only, whereas the condition (8.9) will
usually be applied to electric rather than magnetic effects. The third criterion,
equation (8.6a), resulting from the comparison of the imposed current
density, defines the scale for the magnetic field strength mH on the basis of
current excitation. Following the above procedure it is also possible to define
scaling criteria of processes and physical effects described by other types of
equations.
8.2.2 ModelUng fields and losses in fe"omagnetics

In the majority of cases, various iron components of electrical devices, when
subjected to a harmonic magnetic field, maintain the properties of solid iron.
Moreover, it is often possible to neglect the effect of eddy-currents on the
excitation flux. From a practical point of view, it is usually sufficient to apply
the plane wave theory in order to calculate the local power loss [8.24; 8.29].
As shown in Fig. 8.1 the penetration of flux into solid iron is governed by the
normal component of flux density.
The leakage flux per unit length along the z-axis in the cross-section of iron
at position Y equals
Y
<l>m(Y) = J IBn(y) I dy (8.10)

Yo
Q} b}
A-A
y
...,tDJ.::~~dy
Y Yo
Fig. 8.1 Penetration of magnetic field into solid iron [8.29]: (a) iron plate near current
leads, (b) cross-section
where Yo is the coordinate at which H mt (y) = O. The same flux from the solid
iron side may be expressed as
(8.11)
where:
IJ.m static magnetic permeability IJ.m(Hmt );
'II a semi-empirical coefficient describing an increase of leakage flux due
to the magnetic non-linearity of iron (see reference [8.25]).
Comparison of eqns (8.10) and (8.11) leads to the following relationships:
(8.12)
(8.13)
We adopt, after Neyman [8.16], the approximation:

H = C1B n (8.14)
where C1 is a constant dependent on the actual BIH curve. Following the
author's investigations [8.27; 8.28] we write:
- m(l-n) - m(l-n)/n
m11- (8.15)
B - H
Taking into account eqns (8.7) and (8.15) we find a condition for frequency:
- m(n-I)/n m -l m -2
m1- (8.16)
Hoi
Finally, comparison of eqns (8.12) and (8.13), together with the identity
mw = I, yields:
(8.17)
which implies consistency of modelling of both components of magnetic field
at frequency defined by eqn (8.16).
The flux density is modelled according to the formula:
(8.18)
where:
mj scaling factor for current density;
m n scaling factor for a number of turns.
Thus the frequency is modelled as
ml = (mjmn)n-l m;l m7- 3 (8.19)
The local power loss, in W m -2, will satisfy
mps = (mjm n)2n (m/)2n-l (8.20)

mo
..
-- ,...-- --- ?
(\~\
~ ~
10 n::1
-
V
/ ~
....
V n=3
~
""
............. r--. n=4
.............
-r-- ~5
- ....
2 3 4 5 6 7 8
Fig. 8.2 Frequency scale mf as a function of linear dimensions [8.29) (mj = mn =m a = 1)
~
\\
, --
.~ ~ ,
r-- ~
--- -- -- -
"
.~ ..........
~ n"'1
~ " t".......
'-
- --
~ ..............
~~ ' .... -........ !!.:l.
"
........... ~
~ ' ..... :'2....2
~ .......
~ I)",,)
"'~".....
I ~ ~!}"J.
~
2 3 4 5 6 7 8
"""""
Fig. 8.3 Local power loss scale mps as a function of linear dimensions at frequency given
by eqn (8.19) - continuous line, and at the frequency scale mf= 1 - dashed line [8.29}
(mj = m n = m a = 1)
Coefficients mf and mps are plotted in Figs 8.2 and 8.3 respectively, and
can be used to design reduced-scale models.
A typical electrical power device contains elements of different levels of
magnetic saturation. Therefore, it will not be possible to fulfil the frequency
requirement for the entire object. Modelling can be carried out as a function
of frequency; however, in a given experiment only one frequency of the

external field may be used at a time.
The general formula for the power loss scale mps is as follows:
mps = ~ [mfJ
m m(3n+
Bn
I )/2
a
For a particular case of mf = I, Le. the same frequency in the model and the
original system, the variation of the coefficient mps is demonstrated with a
dashed line in Fig. 8.3.
To determine the maximum allowed modelling frequency at which the
reaction of eddy-current may still be neglected, numerical calculations were
made using a simplified model shown in Fig. 8.4 with the additional assump-
tion that mil = ma = 1[8.31].
The magnetic field strength should obey mH = ml' because mj = mn =
mil = 1. The calculated scaling factor for tangential and normal components
are mHI and mHn respectively. As long as the eddy-current reaction on the
excitation field is not observed, the coefficients WI = mH/mH and Wn =
mHn/mH have values close to unity.
It follows from Fig.8.5, that for a non-screened steel, the maximum
admissible frequency is equal about 2000 Hz.
8.2.3 Linear motors

A significant contribution in the area of electromagnetic field modelling in
rotating machines was made by Ivanov-Smolensky [8.10]. In the field of linear
ds
I 11 1I1 IV
Q1 0,1 0,17 0,1 0,17 m
Q2 0,22 0,29 0,22 0,29 m
Fig. 8.4 Dimensions of the transformer system used to calculate maximum modelling
frequency [8.31}
/11
/
II' III
~Wn
f---Wt I "7 Il-
111_
/ J II /
I
V I ~I
v
-- -- ~ ,- llllV
,
0,5
- -~ -- 1--- 11
50 100 1000 10000 100000

f(Hz)
Fig. 8.5 Coefficients WI and W n for the system from Fig. 8.4 as functions of frequency
[8.31J
motors some interesting papers were published by Bland, Lowther, Freeman

and Laithwaite [8.1, 8.5]. Further studies which take account of magnetic
non-linearity of solid iron are reported in [8.1, 8.5]. In particular the scaling
factor for the magnetomotive force is
2
_ ~--'--"":........;.---"'----'"
mm/- mjm,mk"m,mnmk
(8.21)
mg
where:
mk" scale for the magnetizing current;
m, scale for the coefficient of demagnetizing armature reaction;
mn scale for the number of winding turns;
mk scale for the winding coefficient;
mg scale for the number of parallel paths.
The scaling factors for a magnetic field strength in laminated core, iron
armature and yoke are:
mHe -m
-
-mH_mm/
Ha - y-- (8.22)
m,
In the air-gap, where for technological reasons m" may be different from m"
we have
I'\..
",'" k- n=1
10
""'-.........
N "-~ ~n=5
"- ~n=2
~ ~n-oo
1
"" I'....
"~
"-~ ~
"
1'... ........
-2
10 0,1 0,2 '0,5 2 "'''"
Fig. 8.6 Frequency scale mf as a function of linear dimensions {S.lS} (m o =mj =mkp =m,
= mn =mg =mk = 1)
(8.23)
Substituting (8.21) and (8.22) into eqn (8.16) gives the frequency condition
_ (mjmkfJ m,mnmk) (n-I )/n
mf - (mg)(n I)/n(m/)(n+I)/n(m ) (8.24)
o
The scaling factor for a linear speed may be found from eqn (8.5) as
_ (mjmkfJm,mnmk) (n-I)/n
(8.25)
mv - (mg)(n-I)/n(mYln(m )
o
The plots of mf and m v are presented in Figs 8.6 and 8.7.

The thrust force of the motor may be calculated from the power of the
travelling field penetrating into the solid armature, and the synchronous linear
speed. This power is calculated by integration of the complex Poynting vector
over one or both sides of the armature surface, depending on whether it is
a single-sided or double-sided motor. Thus
P = RegfEmaH:adS} (8.26)
in the model
mpP = mEamHam;P (8.27)
Hence the scaling factor for a motor thrust force is
(8.28)
which after substitution and rearrangement gives

mv 10 .~-'--'-.....,....--,---r----.
5t-~.:........:....-+---I1---+--+_--1
0,2 f-----lf--+------lf---+-~+---1
0,1 '---'-_.....L..----J'-----1-_.....I.-~
0,1 0,2 0,5 2 5 10 m
l
Fig. 8.7 Linear speed scale m v as a function of linear dimension {S.lS} (m a = mj = mk. =
m, = m n = mg = mk = 1)
m = m(3n+l)/n [mom m mn m J(n+l)/n (8.29)

F I J kll' k
mg
The plots of the scaling factor for the motor thrust mF for non-magnetic
or highly-saturated ferromagnetic armature are presented in Fig. 8.8. Follow-
ing a similar procedure it is possible to define modelling scales for power loss
in the inductor winding, iron loss, leakage reactance, supply voltage and
apparent power of the motor.
8.3 EXPERIMENTAL TECHNIQUES

8.3.1 Experimental models
Experiments are often conducted on special models for the purpose of
verifying computational methods and techniques. Such models are of a special
./ <II
-'
d
/ /' u
III
Y" ~
~ E
E
II I- ~ ./ ..c
......
~/~
Eu-
10 -2
~\
10-4
'--- --
0,1 0,2 0,5 2
Fig. 8.8 Motor thrust scale mF as a function of linear dimensions (m a = mj = mk/l = m, =

m n = mg = mk = 1)
Experimental Techniques 339
type because from the point of view of physical modelling theory they are
actually originals. In this case, for distinction, the name experimental models
will be used.
The level of simplification in experimental models varies. The majority of
these models resemble practical devices such as electrical machines, some will
simply be industrial products as such. They are used to assess the adequacy
and accuracy of mathematical models.
Physical fields are described by both space distribution and integral
parameters. Thus the experimental methods must be able to provide measure-
ments of distributed as well as integral values.
8.3.2 Basic measurements

Basic investigations include experiments on ring or tube samples to define the
relation between the electric and magnetic field strength in metals, especially
in iron. In such experiments the boundary and initial conditions are precisely
defined.
The ring sample is excited by a voltage supply (primary winding I of nl
turns), whereas the tube sample by a current flow (see Fig. 8.9). The shape
of a sample enables the calculation of magnetic field strength from simple
analytical formulae. The time wave-forms of currents can be displayed on an
oscilloscope. Electric field strength in the ring sample is found from an
electromotive force induced in the secondary winding 2 of nz turns or in
the case of a tube, by point measurement of electric potential drop along
the current flow. The application of compensation techniques is necessary to
measure accurately such small potentials [8.22; 8.26].
The magnetic field strength, averaged for both surfaces of the ring sample,
is calculated from the formula:
inl
H=- (8.30)
'ltD
Fig. 8.9 Samples used for basic measurements: (a) section through a ring sample
(1 - exciting winding, 2 - voltage winding); (b) tube sample (AB - points of potential
measurements)
in which
D_ ..J2DoD;
(8.31)
- '!(DS + Df)
The electric field strength equals
E= V2 (8.32)
2(h + a)n2
where a and h are shown in Fig. 8.9.
For the tube sample
H =...!.-
1tD
(8.33)
and
(8.34)
The power loss resulting from the Poynting vector may be measured using a
wattmeter or from the oscilloscope by multiplying E by H and integrating over
the surface.
For sinusoidal time variation:
(8.35)
The loss in a sector of the ring sample of length I, according to the field
penetration model, may be written as
(8.36)
The first term in eqn (8.36) gives the rms voltage per turn en whereas the
second yields the mmf drop (magnetic potential difference) VI!; Therefore
the formula for power loss in the sample reads
P = e/ VI! cos q> (8.37)
where q> is the phase angle between the magnetic potential difference and the
induced voltage. If the oscilloscope is used then the power loss is proportional
to the surface of Lissajou figure. In practice, the total power loss is usually
measured using a wattmeter and then divided by the surface of the sample,
to obtain power loss per unit area, which gives a correct result providing the
sample is homogeneous and isotropic. In the ring sample, the excitation
current and induced voltage U2 are used to measure the power. In the case
of the tube, the conduction current and the tube terminal voltage are used
[8.22; 8.26J.
8.3.3 Flux and power loss in magnetic cores

Ferromagnetic cores of transformers, rotating machines, reactors and other
electrical devices are made from laminated electrical steel. The cold-rolled
grain-oriented sheet and strip materials have magnetic properties and power
loss in the rolling direction superior to those on any other axis. Such sheets
have been used in transformers for over 30 years and have also been recently
introduced to rotating machines.
The Epstein's apparatus is the basic instrument enabling measurements of
average unit loss in samples for comparison of different types of magnetic
sheets. However, the measurements tend not to be very accurate and are not
particularly helpful in explaining physical phenomena.
A modified experiment uses single sheets as shown in Fig. 8.10. The
advantage is that -we can define exactly the conditions of field excitation.
This overcomes the problem inherent to the standard Epstein's instrument of
non-uniformity of flux distribution in the packet of laminations and the
anisotropy effect of the corners [8.26].
In Fig. 8.10 the so called Rogowski's coil is presented. The measuring
coil is wound on an elastic non-magnetic core. The flux linking the coil is
given by
\II = l1oSn ' UIJ.AB (8.38)
where:
n' number of turns per coil unit length;
s cross-section area of the coil;
UIJ.AB magnetic potential difference (mmf).
The flux linkage in an alternating field is obtained from the electromotive
force e induced in the coil:
6
H
7
•
Fig. 8.10 Modified Epstein's apparatus for measurements on single sheets: 1 - exciting
winding, 2 - sample (single sheet), 3 - Rogowski's coil, 4 - voltage winding, 5 - core,
6 - integrator, 7 - oscilloscope
Fig. 8.11 Coils for measuring flux in packets of laminations: 1, 2 - packets of lamination,
3, 4 - multiwire voltage probes
e
\11=- (8.39)
4ot!
where Ok is the form coefficient of the electromotive force.
In the magnetostatic field
(8.40)
where:
c'II constant of the ballistic galvanometer;
(l galvanometer indication.
The electric field strength is measured using the winding number 4 shown
in Fig. 8.10. Using a two-beam oscilloscope we can record simultaneously both
the magnetic field strength H and the electric field strength E.
The packets of laminations forming a core may have different reluctances.
Thus it is necessary to measure the total flux and flux distribution between
individual packets. Tests show that even if the total flux in the core is
sinusoidal, the component fluxes may be distorted [8.13]. Multiwire voltage
pick-up coils enclosing or stitched through the lamination are often used
for such measurements (Fig. 8.11). The application of a vector-meter or an
oscilloscope with an integrator enables the time recording of the average flux
density in the core.
The Hall probe may also be used for measuring power loss distribution
in a transformer core. The magnetic field strength· H must be perpendicular
to the probe plane. The terminals must be connected to the measuring coil
enclosing a part of the core adjacent to the probe. Current is then proportional
to electric field strength on the core surface [8.6; 8.22].
8.3.4 Measurements in air-gaps of electrical machines

Mechanical characteristics of rotating machines and linear motors depend on
the magnetic field distribution in their air-gap. Slots in both stator and rotor
3
Fig. 8.12 Fully-pitched coil for measurements of magnetic flux density in the air-gap of
rotating machines I - multiwire coil, 2 - terminals, 3 - slip-rings, 4 - brushes
generate additional harmoriics of flux density. The distribution of the normal

component of flux density may be registered using fixed or movable Hall's
probes in the air-gap region. Other ways of investigating flux distribution are
search coils fixed to the stator surface, or coils wound on the tooth heads to
measure flux entering the iron core.
One of the oldest methods is the application of a fully-pitched coil located
on the rotor and connected to the slip-rings (Fig. 8.12). The electromotive
force of rotation, induced in the coil, represents the air-gap distribution of
flux density in one pole pitch:
e = Blvn (8.41)
where:
B magnetic flux density;
I axial length of the coil;
v speed of rotation of the coil;
n number of turns.
8.3.5 Power loss in transformer tanks

Using the eddy-current flow analogy it is possible to build a simplified model
for investigating eddy-current in a single-phase transformer tank. As shown
in Fig.8.l3, the model is built from a thick steel plate placed between
excitation coils. The eddy-currents on both surfaces of the middle plate of the
model have similar distribution to eddy-currents on the inner surface of a real
transformer tank. The transformer core is modelled using laminations 5. For
~ I
....
,""='\0-
/'.--- "\ -+ 5
,/
ir 5
ff~ ~
I 0) I
x -~././
------ b
/-;.:_-.:..-,
=lit -:::....-
(\ (0
..... _
-3 -....___
5
....
~a--+-).
_--_
0) I\
... "-
'\
~ ...-'I
2
.., 1
h r ... ' ... )
u",I ~
.....
t:s:!
r
.... -
r- - -~ 3
1I1I11I11111}111111 2
5
Fig. 8.13 Model of a single-phase transformer tank for power loss investigations: 1 - steel
plate tank, 2 - exciting winding, 3 - voltage windings, 4 - additional steel plates for balancing
the field distribution, 5 - laminations (transformer core)
power loss measurements the idea of Epstein's apparatus is used. The voltage
winding 3, which has the same number of turns as the excitation winding 2
and the same length as the steel plate, is connected to the voltage coil of
the wattmeter. The current coil of the wattmeter is supplied by the excitation
current directly or via a current transformer. It is possible, for example, to
investigate the effect of eddy-current screening of the tank by covering the
steel plate with aluminium or copper sheets of different thickness.
Similarly, measurements can be made with magnetic shunts which partially
or fully cover the steel plate, to assess the magnetic screening efficiency.
Measurements of normal and tangential components of magnetic field
strength are carried out by means of simple search coils, placed near the core,
shunts, windings etc. Coils stitched through the steel plate enable measure-
ments of electric field strength on the surface, as well as flux distribution along
the x-axis inside the tank.
8.3.6 Leakage fields

The three-dimensional character of leakage fields of most practical devices
makes experiments laborious. The space configuration of leakage fields can
be investigated by Hall probes moved in space relative to a fixed coordinate
system. An alternative system consists of three search coils fixed to a stand
so that the position of a 'measurement point' can be easily defined. Such
investigations are carried out near the transformer windings, inside and
outside the 'window', and in the coil end region in alternating current machines
(Fig. 8.14).
In the case of large power transformers successful measurements have been
reported for leakage fields in tanks, covers, yokes and shields [8.18; 8.2].
Experiments for Coupled Fields 345
Fig. 8.14 Measurements of the transformer leakage field with probes fixed to a stand:
1 - transformer core, 2 - windings
Flux investigation in the closed volume of a transformer tank, or rotating

machine casing, requires a large number of fixed search coils. However these
fields are usually quite strong, well above the background noise level, which,
in addition, is usually well screened by the metallic cover.
The situation is very different with external leakage fluxes which penetrate
through the cover and are thus present in the neighbourhood of electrical
devices. These fields are very weak and are often disturbed by background
fields of, for example, geomagnetic origin. Measurements are made using
special equipment with geomagnetic field compensation (e.g. Forster's
probes). Usually there is automatic control and microcomputer processing of
results [8.3]. The results are often presented graphically as three-dimensional
field plots [8.9].
8.4 EXPERIMENTS FOR COUPLED FIELDS

8.4.1 Forces and stresses
Electromagnetic force is one of the integral parameters of an electromagnetic
field. As distributed systems, these are fields of space force density, surface
stress or tension. Both the distributed and the integral forces or stresses are
measured with tensometers fixed directly or indirectly to appropriate elements
of the device. The tensometer technique requires sensitive bridge methods and
electronic amplifiers for registration of transients [8.20; 8.21].
Each force-measuring system has an oscillating mechanical part (elastic
mass causing the deformation of the element with attached tensometers), and
an electrical part (bridge, amplifier, oscilloscope), see Fig. 8.15.
Because of negligible non-linearity and omission of mechanical hysteresis,
(a) (b)
- , : : ,...m
Fig. 8.15 fa) Measuring system for the linear-motor thrust force, and (b) the measuring
beam
the static sensitivity is often identified with the amplification coefficient K,

defined as a ratio of the output Y u to the value Fu measured in the steady
state [8.8; 8.19]:
K= Yu . (8.42)
Fu
The following equation describes the system
ji - 2~000Y + oo~y = oo~KF(t) (8.43)
where:
~ = dl2..J(Ksm s) relative damping; (8.44)
000 = ..J(Ks/ms ) angular frequency of oscillations; (8.45)
ms reaction mass;
d damping coefficient;
Ks equivalent rigidity.
The time response to the force step impulse is given by
F{t) =Fu[l- ..J{l ~~2) e-I3Cllotsin{OOTt+2q»]I{t) (8.46)
where:
OOT
q> = arctan A (8.47)
..,000
and OOT is the angular frequency of damped oscillations.

By recording the system's response to the force step input it is thus possible
to evaluate the frequency 000 of free oscillations, the relative damping, and
the frequency OOT of damped oscillations.
Better dynamic performance can be achieved by increasing the free vibration
frequency and increasing the damping. On the other hand, however, such
increases are limited by the sensitivity of the measuring and registering
equipment.
The dynamic error
I
F = YJK~ F{t} "IOOO!O (8.48)
should be investigated carefully in every case for a particular measurement.

Vibration measurements are made using capacitive sensors or piezoelectric
transducers [8.21].
Experiments for Coupled Fields 347
The vibroacoustics problems resulting from electromagnetic fields are a

separate area of scientific research [8.14].
Occasionally, observation of winding vibrations in electrical machines and
transformers involves using high speed cameras (thousands of frames per
second).
8.4.2 Temperature fields

The temperature field, resulting from and coupled with the electromagnetic
field in electrical machines and devices, is usually investigated using contact
methods (thermometers, thermoelements). The connections (welding or other
coupling) of copper-nickel or iron-constantan wires, have temperature
dependent electromotive force characteristics [8.21]. Thermocouples with very
small dimensions favour these sensors in applications not only on open
surfaces but also inside the core or windings. Each thermocouple must be
individually calibrated by measuring the voltage-temperature characteristic.
Full automatic control and numerical processing of results is possible and
quite common.
The measurements of high temperature fields are made indirectly by means
of photo-optical pyrometers, such as those used in metallurgy and electro-
thermics. The accuracy at this compared with the contact methods is
considerably smaller.
Recently an extremely sensitive indirect method based on thermal radiation
has become more popular. Thermovision cameras enable measurements and
presentation of thermal field charts of the external surfaces of electrical
devices.
8.4.3 Thermometric method

The field of power loss may be investigated experimentally by a thermometric
method developed by Turowski [8.12; 8.17; 8.22]. The measurement can be
conducted in a short time without heating the entire object.
The thermal balance equation for an arbitrary point of the metal sheet may
be written as
(8.49)
where:
PI unknown power loss;
c specific heat;
Pm specific mass;
A. specific heat conductivity;
a' representative coefficient of the heat exchange with the environment
by convection and radiation;
9 body temperature;
90 ambient temperature;
t time.
Fig. 8.16 Heating and cooling curves of a homogenous object
Assuming that before switching on, the temperature at each point was the
same, then for the time t = 0 eqn (4.49) simplifies to:
PI = CPm[:OJ
v( 1=0
(8.50)
or
(8.51)
where tan Qo is the initial slope of the temperature curve as shown in Fig.
8.16.
The power loss is then calculated using the initial temperature increase.
Before the measurement can be repeated it is necessary to cool the object. The
power loss can also be determined from the rate of temperature changes
caused by switching on and off. This method is known as the two-tangent
method, where
tan «0 = tan «\ + tan «2 (8.52)
It is found from experience that the two-tangent method is less accurate than
the initial temperature increase method.
Errors associated with the use of eqn (8.50) increase with high values of A.
and «' , as well as being due to the non-linearity of power density distribution
on the surface of the body. Other errors depend on the sensitivity and accuracy
of the measuring and data processing instruments.
8.5 ANALOGUE MODELLING

In the early days of magnetostatic field measurements, iron filings were used
to visualize fields of salient poles or multiwire lines. This method however gave
no information about the values of flux density. Thus additional measure-
ments were taken with a search coil and a ballistic galvanometer. Govorkov
Analogue Modelling 349
[8.7] describes a method of calculating field lines and equipotential lines under
certain conditions, because the two systems of lines are orthogonal and the
integral 1B dl along any equipotential line between two adjacent field lines has
the same value everywhere.
Analogue modelling uses the mathematical similarity of different physical
phenomena. The reluctance network method, presented earlier in this book,
is an example of an analogy between the differential form of Poisson's
equation and the Kirchoff equation in electrical circuits.
The analogue field modeIling as an experimental procedure is carried out
on continuous or discrete field analysers. The oldest analysers, which used
the flow of electric current, were an electrolytic tank and a sheet of semi-
conducting paper. These were intended primarily for investigating electrostatic
and magnetostatic fields in two dimensions. In the case of an electrolytic tank
with a sloping bottom the volume field of axial symmetry could also be
modelled.
From the first Maxwell's equation we know that in the current-free regions
(curl H = 0) the magnetic field can be expressed in terms of a magnetic scalar
potential as Vm:
H = -grad Vm (8.53)
In a homogeneous isotropic medium the magnetic scalar potential obeys the
Laplace's equation:
v2 Vm = 0 (8.54)
The steady current flow field may be described by the equation I = GU,
where:
I current;
G electric conductance;
U potential difference.
In the electrostatic field we have the equation Q = CU,
where:
C capacitance;
Q charge.
In the magnetic field, by analogy, we find:
where:
R.. reluctance;
UI! magnetic potential difference (mmf);
cP magnetic flux.
There is an analogy between current I, charge Q and flux cPo There is also
an analogy between the electric voltage U and the magnetic potential
difference Uw An example of the modelling of a part of a synchronous
Fig. 8.17 Modelling of magnetic field of a synchronous generator with the use of an
electrolytic tank [B. 7]
generator with non-equipotential boundary conditions is presented in Fig.

8.17.
The discrete analysers built as networks of resistors or capacitors are used
to provide two- or three-dimensional solutions of the boundary value
problems. The networks are plane or three-dimensional respectively.
The idea of the method of analogy for the case of an eddy-current problem
will be presented using the diffusion equation.
For a section of a transmission line shown in Fig. 8.18 the following four
equations may be written:
(I) for the nonlinear resistance u,(i) = r(i)i:

aj2
au, (i) . ai a2i
ax2 = c-at + gu,(l) + gl at + c/ at2 (8.55)
-
. ai
-
1+--·
dX
i ur ax
rdx Idx ig
u cdx gdx u+ -fuL dx
ox
.....- - - - - dx -------t~
Fig. 8.18 Elementary section of a transmission line [B.32]

Ana/ague ModeJ/ing 351
(2) for the nonlinear inductance", (i) = /(i)i:

a2i a", (i) , ai a 2", (i)
ax2 = g---at + gn + er at + e~ (8.56)
(3) for the nonlinear capacitance q(u) = e(u)u:

a2u aq(u) au a 2u
ax2 = r ---at + rgu + /gat + /e at 2 ' (8.57)
(4) for the nonlinear conductance ig(u) = g(u)u:

a 2u a 2u
=/T
ai (u) au
ax2 + rig(u) + re at + /e at 2 (8.58)
The modelling of a diffusion equation in one dimension, at some depth x

for a thin layer of thickness dx, may be based on the following four elementary
circuits:
(1) quadrupole r(i), e; / = 0, g = 0:

a 2i au,(i)
-= 2 e-- (8.59)
ax at
(2) quadrupole /(i), g; e = 0, r = 0:
a2i a", (i) (8.60)
ax 2 = g---at
(3) quadrupole r, e(u); / = 0, g = 0:
a 2u aq(u)
ax2 = r---at (8.61)
(4) quadrupole /, g(u); e = 0, r = 0:
a2u = / aig(u) (8.62)

ax 2 at
T'he first circuit is described in reference [8.32] andis shown in Fig. 8.19,
where a variable resistor (varistor) is used as the nonlinear element to model
the magnetic characteristic of iron (see Fig. 8.20).
From the similarity of the circuit equation:
, 2' , dUk_lkdik_lk (863)

'k,k+l - 'k-l,k + 'k-2,k-l = e dik-l,~ ~ •
and the differential-difference field equation:
H(x + Ax, t) - 2H(x, t) + H(x - Ax, t) dB (x, t) dH(x, t)

Ax 2 = a dH(x, t) dt (8.64)
.. Uk-2,k-1
~ .. Uk-1,k
• .. Uk,k+1
~
i k-2,k-1 i k-1,k i k k+1

• • I
~
R(i) R(j) R(il
Fig. 8.19 Simple sections forming an analyser for the diffusion equation
S( T)
1,5
,,-
~
~
1,0
1..... ",,'
/ 112
0,5
I
/
. .'
o
i i
2I ,4 5i 6 7 8
i i i
9 l(mA)
o 4 8 16 20 24 28 32 I H)2 (AIm)
Fig. 8.20 The characteristic of a variable resistor (varistor) - curve 1, and a typical B-H
curve - curve 2 f8.32]
the condition of analogy results in
c = (.:iX)2 a dB(x, t)jdUk_I,k (8.65)

dH(x, t) dik_ 1 k
The magnetic hysteresis effect may also be included by connecting a capa-
citor (selected experimentally) in parallel with the resistor in each section of
the analyser. For two- and three-dimensional fields more elaborate networks
are required but they will follow the same principle and similar equations.
8.6 EXPERIMENTAL ERRORS

Suppose the objective value of a given parameter can be defined in some way.
Then the discrepancy between the result obtained from a measurement and
this objective value of the quantity is called the measurement error. Ideally,
the objective value should be chosen as the actual value. This value, however,
Experimental Errors 353
is not known, and alternative definitions are sought. Very often the objective
value is established computationally, especially if calculation is considered to
be more accurate than other methods. Sometimes the average of several
measurements is taken. It is important to realize that it is not possible to
eliminate experimental errors, we can only try to reduce them or minimize
their effect.
In reference [8.21] the following classification of measurement errors is
suggested.
(1) In view of the relationship between the measured result Wand the
objective value Wo:
• absolute errors (expressed in units of the measured quantity):

~W= W- Wo (8.66)
• relative errors (expressed as percentage):
W- Wo
5w = ·100% (8.67)
Wo
(2) Depending on the type of variation of the measured quantity and proper-
ties of the measurement system:
• static errors occurring during steady-state after transients have died

away,
• dynamic errors under transient conditions.
(3) Depending on the regularity of the occurrence of errors during the multiple
repetition of the measurement under the same conditions:
• systematic errors of constant sign and value, changing according to a

known function with varying measurement conditions,
• stochastic errors, which are non-recurrent even under exactly the same
conditions.
(4) Depending on the physical conditions at the time of experiment (reference

conditions):
• basic errors existing in reference conditions,

• supplementary errors resulting from disturbances in reference
conditions.
References
Chapter 1
1.1 Baldomir D.: Differential forms and electromagnetism in 3-dimensional Euclidean space
R 3, lEE Proc. A, 1986, 133, (3), pp. 139-143.
1.2 Baldomir D. and Hammond P.: Geometrical formulation for variational electromagnetics,
lEE Proc. A, 1990, 137, (6), pp. 321-330.
1.3 Baldomir D. and Hammond P.: Geometrical approach to eddy-current systems, lEE Proc.
A, 1993, 140, (2), pp. 166-172.
1.4 Baldomir D. and Hammond P.: Global geometry of electromagnetic systems, lEE Proc.
A, 1990, 137, (6), pp. 321-330.
1.5 Binns K.J., Lawrenson P.J. and Trowbridge C.W.: The Analytical and Numerical Solution
of Electric and Magnetic Fields, John Wiley & Sons, Chichester, 1992.
1.6 Bozorth R.M.: Ferromagnetism, Van Nostrand, Princeton, USA, 1951, p.476.
1.7 Carpenter c.J.: Comparison of alternative formulations of 3-dimensional magnetic field
and eddy-current problems at power frequencies, Proc. lEE, 1977, 124, (11), pp. 1026-
1034.
1.8 Carpenter C.J.: Electromagnetic energy changes due to charges moving through constant,
or zero, magnetic field, lEE Proc. A, 1991, 138, (I), pp. 55-70.
1.9 Carpenter C.J.: Electromagnetic theory without electric flux, lEE Proc. A, 1992, 139, (4),
pp. 189-209.
1.10 Carter R.G.: Electromagnetism for Electronic Engineers, Chapman & Hall, London, 1992.
1.11 Carter G.W.: The Electromagnetic Field in its Engineering Aspects, Longmans, Green and
Co, London, 1957.
1.12 Chikazumi S.: Physics of Magnetism, Wiley, New York, 1964, p.245.
1.13 Christopoulos C.: An Introduction to Applied Electromagnetism, John Wiley & Sons,
Chichester, 1990.
1.14 Dobbs E.R.: Electricity and Magnetism, Routledge & Kegan Paul, London, 1984.
1.15 Ferrari R.L.: Complementary variational formulation for eddy-current problems using the
field variables E and H directly, lEE Proc. A, 1985, 132, (4), pp. 157-164.
1.16 Hammond P. and Sykulski J.K.: Engineering Electromagnetism: Physical Processes and
Computation, Oxford University Press, Oxford, 1994.
1.17 Hammond P. and Sykulski J.K.: Tubes and Slices: a new way of teaching the principles
of electric and magnetic fields, IEEE Trans. Educ., 1992, 35, (4), pp. 300-306.
1.18 Hammond P.: Electromagnetism for Engineers, Pergamon Press, Oxford, 1978.
1.19 Hammond P.: Applied Electromagnetism, Pergamon Press, Oxford, 1988.
1.20 Hammond P.: Energy Methods in Electromagnetism, Clarendon Press, Oxford, 1981.
1.21 Hammond P.: Use of potentials in calculation of electromagnetic fields, lEE Proc. A, 1982,
129, (2), pp. 106-112.
1.22 Hammond P. and Baldomir D.: Dual energy methods in electromagnetism using tubes and
slices, lEE Proc. A, 1988, 135, (3), pp. 167-172.
1.23 Harrington R.F.: Introduction to Electromagnetic Engineering, McGraw-Hill Book
Company, New York, 1958.
1.24 Hayt W.H.: Engineering Electromagnetics, McGraw-Hili Book Company, 1974.
References 355
1.25 Krawczyk A. and Tegopoulos J.A.: Numerical Modelling of Eddy Currents, Clarendon
Press, Oxford, 1993.
1.26 Lammeraner J. and Stafl M.: Eddy Currents, Iliffe Books Ltd, London, 1966.
1.27 Lancaster G.: Introduction to Fields and Circuits, Oxford University Press, New York,
1992.
1.28 Liao S.Y.: Engineering Applications of Electromagnetic Theory, West Publishing
Company, St Paul, 1988.
1.29 Lorrain P., Corson D.P. and Lorrain F.: Electromagnetic Fields and Waves, W.H. Freeman
and Company, New York, 1988.
1.30 Maxwell J.C.: A Treatise on Electricity and Magnetism, Clarendon Press, Oxford, 1892.
1.31 Moon P. and Spencer D.E.: Field Theory for Engineers, D. Van Nostrand Company, New
York, 1961.
1.32 Penman J. and Fraser J.R.: Unified approach to problems in electromagnetism, lEE Proc.
A, 1984, 131, (I), pp. 55-61.
1.33 Penman J. and Fraser J.R.: Dual and complementary methods in electromagnetism, IEEE
Trans., 1983, MAG-19, pp.2311-2316.
1.34 Popovic B.D.: Introductory Engineering Electromagnetics, Addison-Wesley, 1971, pp. 416-
418.
1.35 Rikabi J., Bryant C.F. and Freeman E.M.: Error-based derivation of complementary
formulations for the eddy-current problem, lEE Proc. A, 1988, 135, (4), pp.208-216.
1.36 Smythe W.R.: Static and Dynamic Electricity, 2nd edn, McGraw-Hill, New York, 1950.
1.37 Solymar L. and Walsh D.: Lectures on the Electrical Properties of Matters, 4th edn, Oxford
University Press, 1988.
1.38 Spiegel M.R.: Theory and Problems of Vector Analysis, Schaum Publishing Company, New
York, 1959, p. 121.
1.39 Stoll R.L.: The Analysis of Eddy Currents, Clarendon Press, Oxford, 1974.
1.40 Stoll R.L., Bodner B. and AI-Khoury A.H.N.: Method of magnetic sources applied to
the basic configuration of a rotating electrical machine, lEE Proc. D, 1993, 140, (3),
pp.217-231.
1.41 Stratton J.A.: Electromagnetic Theory, McGraw-Hill, New York, 1941.
1.42 Sykulski J.K. and Turowski J.: ISEF'91 Proceedings, COMPEL, Vol. II, No. I, James &
James Science Publishers, London, March 1992.
1.43 Sykulski J .K.: Computer package for calculating electric and magnetic fields exploiting dual
energy bounds, lEE Proc. A, 1988, 135, (3), pp. 145-150.
1.44 Sykulski J .K.: Tubes and Slices: a software package for dual field calculations, in
Software Applications in Electrical Engineering, Computational Mechanics Publications,
Southampton, 1993, pp. 247-254.
Cbapter 2
2.1 'Abrams M.D. and Gillott D.H.: Numerical analysis of hysteresis and eddy current losses
in solid cylindrical rods of No 1010 steel, IEEE Trans., 1967, PAS·86, (9), pp. 1077-1082.
2.2 Ames W.F.: Numerical Methodsfor Partial Differential Equations, Nelson, London, 1969.
2.3 Binns K.J. and Lawrenson P.J.: Analysis and Computation of Electric and Magnetic Field
Problems, Pergamon Press, Oxford, 1963.
2.4 Binns K.J., Lawrenson P.J. and Towbridge C.W.: Analytical and Numerical Solution of
Electric and Magnetic Field Problems, John Wiley, Chichester, 1992.
2.5 Burais N. and Grellat G.: Numerical modelling of iron losses in ferromagnetic steel plate,
IEEE Trans., 1982, MAG-18, pp. 558-562.
2.6 Buzbee B., Dorr F., George J. and Golub G.: The discrete solution of Poisson equation
on irregular regions, SIAM J. Numer. Anal., 1971, 8(4), pp. 722-736.
2.7 Carpenter C.J.: Comparison of alternative formulations of 3-dimensional magnetic-field
and eddy-current problems at power frequencies, Proc. lEE, 1977, 124, (II), pp. 1026-
1034.
356 References
2.8 Carpenter C.J. and Wyatt E.A.: Efficiency of Numerical Techniquesfor Computing Eddy
currents in Two and Three dimensions, COMPUMAG, Oxford, 1976, pp.242-250.
2.9 Carre B.A.: The determination of the optimum accelerating factor for successive over-
relaxation, Computer Journal, 1961, 4, pp. 73-78.
2.10 Coulson M.A., Slater R.D. and Simpson R.R.S.: Representation of magnetic characteristic,
including hysteresis, using Priesach's theory, Proc. lEE, 1977, 124, (10), pp. 895-898.
2.11 Conte S.D. and De Boor C.: Elementary Numerical Analysis, McGraw-Hill Kogakusha,
Tokyo, 1972.
2.12 Davey K.R. and King E.I.: A 3-dimensional scalar potential field solution and its application
to the turbine generator end region, IEEE Trans., 1981, PAS-loo, (5), pp.2302-231O.
2.13 De la Vallee Poussin F. and Lion A.: Calcul iteratif de I'induction magnetique dans les
machines electriques, Rev. Generale de I'Electricite, April 1967, pp.731-739.
2.14 Del Vecchio R.M.: An efficient procedure for modelling complex hysteresis processes in
ferromagnetic materials, IEEE Trans., 1980, MAG-16, (5), pp.809-811.
2.15 DuFort E.C. and Frankel S.P.: Stability conditions in the numerical treatment of parabolic
differential equations, Mathl. Tabl. Natn. Res. Coun., Washington, 1953,7, pp. 135-152.
2.16 Erdelyi E.A. and Fuchs E.F.: Nonlinear magnetic field analysis of DC machines. Part I:
Theoretical fundamentals, IEEE Trans., 1970, PAS-89, (7), pp. 1546-1554.
2.17 Fischer J. and Moser H.: The representation of the magnetization curve by simple algebraic
and transcendental functions, Arch. Elektrotech, 1956, 42, pp. 286-299.
2.18 Fuchs E.F. and Erdelyi E.A.: Nonlinear magnetic field analysis of DC machines. Part II:
Application of the improved treatment, IEEE Trans., 1970, PAS-89, (7), pp. 1555-1564.
2.19 Garadedian P.R.: Estimation of the relaxation factor for small mesh size, Math. Tabl. Natn.
Res. Coun., Washington, 1956, 10, pp. 183-185.
2.20 Gourlay A.R.: Hopscotch: A fast second-order partial differential equation solver, J. Inst.
Math. Appl., 1970, 6, pp. 375-390.
2.21 Hockney R.W.: A fast direct solution of Poisson's equation using Fourier analysis, J. Assoc.
Comput. Mach., 1965, 12, (I), pp.95-113.
2.22 Jack A.G. and Stoll R.L.: Negative-sequence currents and losses in the solid rotor of a
turbogenerator, lEE Proc. C, 1980, 127, (2), pp. 53-64.
2.23 Kozakoff D.J. and Simons F.O.: Three-dimensional nonlinear magnetic field boundary
value problem and its numerical solution, IEEE Trans., 1970, MAG-6, (4), pp. 828-833.
2.24 Kraus J.D.: Electromagnetics, McGraw-Hill, New York, 1991.
2.25 Krawczyk A. and Tegopoulos J.A.: Numerical Modelling of Eddy Currents, Clarendon
2.26 Kronsjo L: Algorithms; their Complexity and Efficiency, John Wiley, Chichester, 1979.
2.27 Leurs L. and Stoll R.L.: Three-dimensional quasi-static magnetic field in superconducting-
rotor synchronous generators with a magnetic steel rotor screen, lEE Proc. C, 1986,133,
(2) pp. 69-80.
2.28 Lim K.K. and Hammond, P.: Universal loss chart for the calculation of eddy-current losses
in thick steel plates, Proc. lEE, 1970, 117, (4), pp. 857-864.
2.29 Lim K.K. and Hammond, P.: Numerical method for determining the electromagnetic field
in saturated steel plates, Proc. lEE, 1972, 119, (II), pp. 1667-1674.
2.30 Lukaniszyn M.: Analiza pola elektromagnetycznego przekladnikow pradowych 0 rdzeniach
ramkowych, Politechnika Lodzka, Praca doktorska, Lodz 1985.
2.31 Peaceman D.W. and Rachford H.H;: The numerical solution of parabolic and elliptic
differential equations, J. Soc. Ind. Appl. Math., 1955, 3, pp. 28-41.
2.32 Poritsky H. and Butler J.M.: A.C. flux penetration into magnetic materials with saturation,
IEEE Trans., 1964, COM-83, pp.99-1I1.
2.33 Proskurowski W. and Widlund 0.: On the numerical solution of Helmholtz's equation by
the capacitance matrix method, Math. Comput., 1976, 30, (135), pp.433-468.
2.34 Rhyner J.: Magnetic properties and AC losses of superconductors with power law current-
voltage characteristics, Physico C, 1993, 212, pp.292-3oo.
2.35 Richtmyer R.D. and Morton K.W.: Difference Methods for Initial Value Problems, John
Wiley-Interscience, New York, 1967.
References 357
2.36 Rivas J.: Simple approximations for magnetization curves and hysteresis loops, IEEE
Trans., 1981, MAG-17, (4), pp. 1498-1502.
2.37 Sarma M.S.: Potential functions in electromagnetic field problems, IEEE Trans., 1970,
MAG-6, (3), pp.513-518.
2.38 Saulyev V.K.: Integration of Equations of Parabolic Type by the Method of Nets,
Pergamon Press, Oxford, 1964.
2.39 Schumann U. (ed.): Computers fast elliptic solvers and applications, Proc. of the GAMM-
Workshop on Fast Solution Methods for the Discretized Poisson Equation, Karlsruhe,
1977, Advance Publ., London 1978.
2.40 Schumann U. and Sweet R.: A direct method for solution of Poisson's equation with
Neumann boundary conditions on a staggered grid of arbitrary size, J. Comput. Phys.,
1976, 20, (2), pp. 171-182.
2.41 Smith G.D.: Numerical Solution of Partial Differential Equations: Finite Difference
Methods, Clarendon Press, Oxford, 1978.
2.42 Stoll R.L.: Solution of linear steady-state eddy-current problems by complex successive
overrelaxation, Proc. lEE, 1970, 117, (7), pp. 1317-1323.
2.43 Stoll R.L.: The Analysis of Eddy Currents, Clarendon Press, Oxford, 1974.
2.44 Stoll R.L.: Modelling very small airgaps between magnetic surfaces in finite-difference field
solutions, COMPEL, 1992, 11, (1), pp.77-80.
2.45 Stoll R.L. and Hindmarsh R.: Reduction of local eddy-current losses in turbogenerator
stator windings using composite subconductors, Proc. lEE, 1976, 123, (11), pp. 1217-
1222.
2.46 Stone H.L.: Iterative solution of implicit approximations of multidimensional partial
differential equations, SIAM J. on Numer. Anal., 1968, S, (3), pp.530-558.
2.47 Tegopoulos J.A. and Kriezis E.E.: Eddy Currents in Linear Conducting Media, Elsevier,
Amsterdam, 1985.
2.48 Temperton C.: Direct methods for the solution of the discrete Poisson equation: some
comparisons, J. Comput. Phys., 1979, 31, (1), pp. 1-20.
2.49 Trutt F.C., Erdeli E.A. and Hopkins R.E.: Representation of the magnetic characteristics
of dc machines for computer use, IEEE Trans., 1968, PAS-87, pp.665-669.
2.50 Varga R.S.: Matrix Iterative Analysis, Prentice-Hall, New Jersey, 1962.
2.51 Von Rosenberg D.U.: Methods for the Numerical Solution of Partial Differential
Equations, Elsevier, New York, 1969.
2.52 Wachspress, E.L.: Iterative Solution of Elliptic Systems and Applications to the Neutron
Diffusion Equations of Reactor Physics, Prentice-Hall, New Jersey, 1966.
2.53 Wiak S.: Method of calculation of the transient electromagnetic field in ferromagnetic
materials for the known vector of induction B on the surface, lEE Proc. A, 1984, 131,
(5), pp. 293-300.
2.54 Wiak S.: A method of calculation of transient electromagnetic field in non-linear ferro-
magnetic medium, Arch. Elektrotech., 1985, 68, pp. 113-120.
2.55 Wiak S.: Stability, convergence and accuracy of the DuFort and Frankel difference diagram
approximating a class of non-linear parabolic field equations, Int. J. Numer. Methods
Eng., 1987, 24, pp. 1421-1437.
2.56 Wiak S.: Analysis of difference diagrams stability approximating a class of one- and two-
dimensional non-linear parabolic field equations, Int. J. Numer. Methods Eng., 1989,
28, pp. 1995-2020.
2.57 Wiak S. and Pelikant A.: Difference method in integral technique in field analysis of doubly
salient switched reluctance motor, Arch. Elektrotech., 1989, 72, pp. 381-387.
2.58 Widger G.F.T.: Representation of magnetization curves over extensive range by rational-
fraction approximations, Proc. lEE, 1969, 116, (1), pp. 156-160.
2.59 Wood, W.L.: Practical Time-Stepping Scheme, Clarendon Press, Oxford, 1990.
2.60 Zakrzewski K. and Lukaniszyn M.: Three-dimensional model of three-phase transformer
for leakage field calculation, Arch. Elektrotech., 1990, 73, pp.319-324.
2.61 Zakrzewski K. and Lukaniszyn M.: Three-dimension model of one- and three-phase trans-
former for leakage field calculation, IEEE Trans., 1992, MAG·28, (2), pp. 1344-1347.
358 References
2.62 Zakrzewski K. and Pietras F.: Method of calculating the electromagnetic field and power
losses in ferromagnetic materials, taking into account magnetic hysteresis, Proc. lEE,
1971, 118, (11), pp. 1679-1685.
Cbapter 3
3.1 Alarcon E., Reverter A. and Molina J.: Hierarchical boundary elements, Computer and
structures, 1985, 20, (1-3), pp. 151-156.
3.2 Babuska T.: Trends in finite elements, IEEE Trans. on Mag., 1989, 2S (4).
3.3 Bettess P.: Infinite elements, Int. 1. Num. Methods Eng., 1977, 11, pp.53-64.
3.5 Brebbia C.A.: The Boundary Element Method for Engineers, Pentech. Press, London,
1980.
3.6 Brebbia C.A.: Boundary Element Methods in Engineering, Springer-Verlag, Berlin, 1982.
3.7 Chari M.V.K.: Non linear magnetic field analysis of d.c. machines, IEEE Trans., 1970,
PAS·89, (7), pp.85-89.
3.8 Collatz L.: Numerischel Behandlung von Differentialgleichungen, Springer-Verlag, Berlin,
1955.
3.9 Collins R.J.: Bandwidth reduction by automatic renumbering, Int. 1. Num. Methods Eng.,
1973, 6, pp. 345-356.
3.10 Coulson M.A., Preston T.W. and Reece A.B.J.: 3-dimensional finite-element solvers for
the design of electrical equipment, Compumag, 1985, Fort Collins, USA. (GEC Power
Transformers, Stafford, UK).
3.11 Dabrowski M.: Electrical Machines Magnetic Fields and Circuits, PWN, Warsaw, 1971 (in
Polish).
3.12 Douglas J.Jr. and Dupont Z.: Galerkin methods for parabolic equation, SIAM 1. Numer.
Anal., 1970, 7, (4), pp. 575-626.
3.13 Enderle G., Kansy K. and Pfaff G.: Computer Graphic Programming, Springer Verlag,
Berlin, 1984.
3.14 Feirweather G.: Finite Element Galerkin for Differential Equations, Marcel Dekker, NY,
1978.
3.15 Finlayson B.A.: The Method of Weighted Residuals and Karietional Principles, Academic
Press, NY, 1972.
3.16 Foley J. and van Dam A.: Fundamentals of Interactive Computer Graphics, Addison
Wesley, London, 1982.
3.17 Gramz M. and Zi6tkowski M.: Calculation of Three-Dimensional Electromagnetic Fields,
TU of Szczecin, 1991 (in Polish).
3.18 Gratkowski S.: More on a simple infinite element, Int. 1. for Compo Matut. in EI. and Elec.
Eng., COMPEL, 1986, S, (191-194).
3.19 Gratkowski S. and Zi6kowski M.: A three-dimensional infinite element for modelling open
boundary field problems, IEEE Trans. Magn., 1992, 28, pp. 1675-1678.
3.20 Janke, Emde, Losch.: Tafeln Hoherer Funktionen, B.G. Teubner Verlagsgesellsehaft, Stut-
tgart, 1960.
3.21 Jennings A.: Matrix Computation for Engineers and Scientists, J. Wiley, London, 1977.
3.22 Kasper M. and Kost A.: Finite element field analysis in ferromagnetic materials under
consideration of anisotropy and saturation; Electrosoft, 1990, I, (3).
3.23 Komeza K. and Krusz G.: Hermitian finite elements in numerical analysis of transformer
leakage magnetic field, Proc. Internat. Symposium on Electromagn. Fields in Electrical
Eng., ISEF' 89, 1989, pp. 109-112.
3.24 Macker1e J. and Brebbia C.A.: The Boundary Element Reference Book, Springer-Verlag,
Berlin, 1988.
3.25 Nakata T.: 3-D electromagnetic field analysis, COMPEL, 9, Sup. A James and James, 1990.
3.26 Norris D.H. and de Vries G.: An Introduction to Finite Element, Academic Press, NY, 1978.
3.27 Pissanetzky S.: A simple infinite element, COMPEL, 1984, 3, pp. 107-114.
References 359
3.28 Ratnajeevan S. and Hoole S.: Computer-Aided Analysis and Design of Electromagnetic
Devices, Elsevier, New York, 1989.
3.29 Sikora R.T.: Electromagnetic field theory, WNT, Warsaw, 1985 (in Polish).
3.30 Silvester P.P. and Chari M.V.K.: Finite element solution of saturable for potential
problems, IEEE Trans., 1970, PAS·89, (7), pp. 1642-1951.
3.31 Silvester P. and Ferrari R.: Finite Elements for Electrical Engineers, Cambridge University
Press, London, 1983.
3.32 Strany G. and Fix G.J.: An Analysis of the Finite Element Method, Prentice-Hall, 1973.
3.33 Tandon S.C., Armor A.F. and Chari M.L.K.: Nonlinear transient finite element field
computation for electrical machines and devices, IEEE Trans., 1983, PAS·I02, (5),
pp. 1089-1096.
3.34 Turowski J., Pawluk K., Sikora R. and Zakrzewski K.: Elektromagnetic Field Analysis and
Synthesis, Polish Academy of Science, 'Ossolineum' Wrocaw 1990 (in Polish).
3.35 Walker C. and Brebbia C.A.: Boundary Element Technique in Engineering, Butterworth,
London, 1980.
3.36 Yuan J.S. and Fitzsimons C.J.: A Mesh Generator for Tetrahedral Elements using Delaunay
Triangulation, Rutheford Appleton Laboratory, Sept. 1992.
3.37 Young D.M.: Iterative Solution of Large Linear Systems, Academic Press, NY, 1971.
3.38 Zienkiewicz O.C.: The Finite Element Method, McGraw-Hili, London, 1977.
3.39 Zienkiewicz O.C., Kelly D.W., Gago J. and Babuska I.: Hierarchical finite element
approaches, error estimates and adaptive refinement, Proc. 3rd Congress on Finite
Element Methods, Los Angeles, 1981.
3.40 Akin J.E.: Application and Implementation of Finite Element Methods, Academic Press,
London, 1982.
3.41 Hannakam L. and Albach M.: Losung des Randwertproblems bei Anwesenheit leitender
Korper hoher Permeabilitiit im erregenden magenetischen Wechselseld, Archiv fur
Elektrotechnik, 65, 1982, pp. 117-122 (in German).
3.42 Albach M. and Hannakam L.: Bestimmung induzierter Wirbelstrome in Korpern hoher
Permeabilitiit nach dem Differenzenverfahren, Archiv fur Elektrotechnik, 66, 1983,
pp.67-74 (in German).
3.43 Forghani B., Freeman E.M., Lowther D.A. and Silvester P.P.: Interactive modelling of
magnetization curves, IEEE Trans. Mag-18, 1982, pp. 1070-1072.
3.44 Nakata T. and Kawas Y.: Numerical analysis of nonlinear transient magnetic field using
the finite element method, Electrical Engineering in Japan, 1984, 104, (4), pp.81-87.
3.45 Zi6l'kowski M.: Electromagnetic field analysis in non linear conducting media, PhD thesis,
1984, TU of Szczecin (in Polish).
3.46 Aldefeld B.: Electromagnetic field diffusion in ferromagnetic materials, Proc. lEE, 1978,
125, (4), pp. 278-282.
3.47 Aldefeld B.: Numerical calculation of electromagnetic actuators, Archiv fur Electrotechnik,
64, 1979, pp. 347-352.
3.48 Krawczyk, A. and Tegopoulos J.A.: Numerical Modelling of Eddy Currents, Oxford
Science Publications, 1993.
3.49 Lim K.K. and Hammond P.: Numerical method for determining the electromagnetic field
in saturated steel plates, Proc. lEE, 1972, 119, (11), pp. 1667-1674.
3.50 Lavers J.D.: Finite element solution of nonlinear two dimensional TE-mode eddy current
problems, Trans. IEEE, 1983, Mag-19, (5), pp. 2201-2203.
3.51 Luetke-Daldrup B.: Comparison of exact and approximate finite-element solution of the
two-dimensional nonlinear eddy-current problem with measurements, INTERMAG 1984,
Hamburg.
3.52 Chari M.V.K.: Finite element solution of eddy-current problem in magnetic structures,
IEEE Trans, 1974, PAS-93, (I), pp.62-72.
3.53 Foggia A., Sabonnadiere J.C. and Slivester P.: Finite element solution of saturated
travelling magnetic field problems, IEEE Trans, 1975, PAS·94, (3), pp.866-871.
3.54 Yamada S., Kanamaru Y. and Bessho K.: The transient megnetization process and opera-
tion in the plunger type electromagnet, IEEE Trans, 1976, Mag-12, (6), pp. 1056-1058.
360 References
Chapter 4
4.1 Allan D.J. and Harrison T.H.: Design for reliability of high-voltage power transformers
and reactors, GEC Rev., 1985, I, (1), pp.44-50.
4.2 Anuszczyk J.: Rotational magnetization in magnetic circuits of electrical machines, Proc.
ISEF' 89, (eds: J. Turowski and K. Zakrzewski), James & James Science Publishers,
London, 1990, pp. 147-150.
4.3 Carpenter C.J.: Magnetic equivalent circuits, Proc. lEE, 1968, 115, (10), pp. 1503-1512.
4.4 Carpenter C.J.: Finite-element network models and their application to eddy-current
problems, Proc. lEE, 1975, 122, (4), pp.455-462.
4.5 Chua L.O. and Lin P.M.: Computer-Aided Analysis of Electronic Circuits: Algorithms and
Computational Techniques, Prentice-Hall, Englewood Cliffs, 1975.
4.6 Csendes Z.J., Freeman E.M., Lowther D.A. and Silvester P.P.: Interactive computer
graphics in magnetic field analysis and electric machine design, IEEE Winter Power
Conference, 1980.
4.7 Davey K.R. and King E.J.: A three dimensional scalar potential field solution and its
application to turbine generator end region,lEEE Trans., 1981, PAS·lOO, (5), pp. 2302-
2310.
4.8 Demenko A.: Equivalent RC networks with mutual capacitances for electromagnetic
field simulation of electrical machine transients, IEEE Trans. Magn., 1992, 28, (2),
pp. 1406-1409.
4.9 Djuroviz M. and Carpenter C.J.: 3-dimensional computation of transformer leakage fields
and associated losses, IEEE Trans. Magn., 1975, MAG·11, (5), pp. 1535-1537.
4.10 Djuroviz M. and Monson J.E.: Stray losses in the step of a transformer yoke with a
horizontal magnetic shunt, IEEE Trans., 1982, PAS·IOI, (8), pp.2995-3ooo.
4.11 Eisenstat S., Gursky M., Schultz M. and Sherman A.: Yale Sparse Matrix Package, Dept
of Compo Science, Yale Univ., New Haven, CT, Techn. Report 114, 1977.
4.12 Hammond P.: Applied Electromagnetism, Pergamon Press, Oxford, 1979.
4.13 Hammond P. and Sykulski J.K.: Tubes and slices: a new way of teaching the principles
of electric and magnetic fields, IEEE Trans. Educ., 1992, 35, (4), pp. 300-306.
4.14 Ho C.W., Ruehli A.E. and Brennan P.A.: The modified nodal approach to network
analysis, IEEE Trans. Circuits Syst., 1975, CAS·22, (6), pp. 504-509.
4.15 King E.I.: Equivalent circuits for two-dimensional magnetic fields. I. The static field. II.
The sinusoidally time-varying field, IEEE Trans., 1966, PAS-85, (9), pp.927-945.
4.16 Komeza K., Krusz G. and Turowski J.: Comparison of network and finite element approach
to the solution of stray problems, Proc. ICEM' 84, Part I, Lausanne, 1984, pp. 17-
19.
4.17 Kubusch N.: Berechnung der Verlustdichtverteilung und der Temperaturfelder inaktiver
Konstruktionsteile von Transformatoren, Proc. Int. Symp. Electrodynamic, Forces and
Losses in Transformers, Lodz, 1979, pp.84-95.
4.18 Oberred K.: Evaluation of magnetic fields, eddy-currents and forces in complicated cases
by simulation on grid models (in German), A rchiv fur Elektrotechnik, 1963, XLVIII, (5),
pp.297-313.
4.19 Laithwaite E.R.: Magnetic equivalent circuits for electrical machines, Proc.IEE, 1967,114,
(11), pp. 1805-1809.
4.20 Rais V.R., Turowski J. and Turowski M.: Reluctance network analysis of coupled fields
in reversible electromagnetic pump, Proc. ISEF' 87, Plenum Press, New York, 1988,
pp. 279-283.
4.21 Tinney W.F., Brandwajn V. and Chan S.M.: Sparse vector methods, IEEE Trans., 1985,
PAS·I04, (2), pp.295-301.
4.22 Turowski J.: Technical Electrodynamics, WNT, Warsaw, 1993 (in Polish).
4.23 Turowski J.: Electromagnetic Calculation of Components of Electrical Machines and
Devices, WNT, Warsaw, 1982 (in Polish).
4.24 Turowski J. and Kopec M.: 3-D hybrid, analytically-numerical computation of tank losses
in 3-phase power transformers, Proc. ISEF'93, Warsaw, 1993, pp. 131-134.
References 361
4.25 Turowski J. and Turowski M.: A network approach to the solution of stray field problem
in large transformers, Rozprawy Elektrotechniczne, Polish Academy of Science, No.2,
1985, pp.405-422
4.26 Turowski J., Turowski M. and Kopec M.: Method of three-dimensional network solution
of leakage field in three-phase transformers, IEEE Trans. Magn., 1990,26, (5), pp. 2911-
2919.
4.27 Turowski J., Zakrzewski K., Sikora R. and Pawluk K.: Analysis and Synthesis of
Electromagnetic Fields, (ed. J. Turowski), Polish Academy of Science, Ossolineum,
Wroclaw, 1990 (in Polish).
4.28 Wiak S., Pelikant A. and Turowski J.: Solution of TEAM Workshop Problem 7 by
Reluctance Network Method, Proc. ISEF'93, Warsaw, 1993, pp. 37-40.
4.29 Worotynski J., Turowski M. and Mendrela E.A.: The accuracy of calculation of magnetic
fields, inductance and forces in electromagnetic devices using the reluctance network
method (RNM), Proc. ISEF'93, Warsaw, 1993, pp. 159-162.
Cbapter 5
5.1 Adamiak K.: Application of coils with current to the magnetic field synthesis, Sci. Electr.,
1981, 27, pp. 15-30.
5.2 Adamiak K.: Method of the magnetic field synthesis on the axis of cylinder solenoid,
J. Appl. Phys., 1978, 16, pp.417-423.
5.3 Adamiak K.: Synthesis of homogeneous magnetic field in internal region of cylindrical
solenoid, Arch. Elektrotech., 1980, 62, pp.75-79.
5.4 Armstrong A., Fan M.W., Simkin J. and Trowbridge C.W.: Automated optimization of
magnet design using the boundary integral method, IEEE Trans. Magn., 1982, MAG-lB,
(2), pp. 620-623.
5.5 Arumugam G., Neittaanmaki P. and Salmenjoki K.: Optimal shape design of an
eletromagnet, ZAMM, 1989, 69, (4), pp.234-237.
5.6 Bellina F., Campostrini P., Chitarin G., Stella A. and Trevisan F.: Automated optimal
design techniques for inverse electromagnetic problems, IEEE Trans. Magn., 1992,
MAG-28, (2), pp. 1549-1552.
5.7 Borghi C.A., Reggiani U. and Zama G.: A method for the solution of an axisymmetric
magnetic field synthesis problem, IEEE Trans. Magn., 1991, MAG-27, (5), pp.4093-
4096.
5.8 Di Barba P., Navarra P., Savini A. and Sikora R.: Optimum design of iron-core electro-
magnets, IEEE Trans. Magn., 1990, MAG-26, (2), pp.646-649.
5.9 Grago G., Manella A., Nervi M., Repetto M. and Secondo G.: A combined strategy for
optimization in non linear magnetic problems using simulated annealing and search
techniques, IEEE Trans. Magn., 1992, MAG·2B, (2), pp. 1541-1544.
5.10 Gitosusastro S., Coulomb J.L. and Sabonadiere J.C.: Performance derivative calculations
and optimization procedures, IEEE Trans. Magn., 1989, MAG-2S, (4), pp. 2834-2389.
5.11 Gottvald A.: Optimal magnet design for NMR, IEEE Trans. Magn., 1991, MAG-26, (2),
. pp. 399-401.
5.12 Gottvald A., Preis K., Magele C., Biro O. and Savini A.: Global optimization methods for
computational electromagnetics, IEEE Trans. Magn., 1992, MAG·2B, (2), pp. 1537-1540.
5.13 Guarnieri M., Stella A. and Trevisan F.: A methodological analysis of different
formulations for solving inverse electromagnetic problems, IEEE Trans. Magn., 1990,
MAG·26, (2), pp.622-625.
5.14 Gyimesi M.: Inverse field calculation for structure positioning, IEEE Trans. Magn., 1991,
MAG-27, (5), pp.4170-4172.
5.15 Hadamard J.: Sur les problemes aux derivees partielles et leur signification physique, Bull.
Univ. Princeton, 1902, p. 13.
5.16 Henneberger G., Meunier G., Sabonnadiere J.C., Settler P. and Shen D.: Sensitivity
analysis of the nodal position in the adaptive refinement of finite element meshes, IEEE
Trans. Magn., 1990, MAG-26, (2), pp.787-79O.
362 References
5.17 ll-han Park, Beom-taek Lee and Song-yop Hahn: Design sensitivity analysis for nonlinear
magnetostatic problems using finite element method, IEEE Trans. Magn., 1992,
MAG-28, (2), pp. 1533-1536.
5.18 Iwamura Y. and Miya K.: Numerical approach to inverse problem of crack shape
recognition based on the electrical potential method, IEEE Trans. Magn., 1990, MAG·26,
(2), pp.618-621.
5.19 Kahler G.R. and Della Torre E.: Minimizing the deformation of a static magnetic field by
the presence of a ferromagnetic body, IEEE Trans. Magn., 1991, MAG·27, (6),
pp.5025-5027.
5.20 Kasper M.: Shape optimization by evolution strategy, IEEE Trans. Magn., 1992, MAG-28,
(2), pp. 1556-1559.
5.21 Korn G.A. and Korn T.M.: Mathematical Handbook, McGraw Hill, New York, Toronto,
London, 1961.
5.22 Krawczyk A.: Application of the boundary element method to transient electromagnetic
problems, Pro Inst. Elektrotech., 1984, 132, pp.85-95.
5.23 Lattes R. and Lions J.: Methode de Quas-Reversibilite et Applications. Dunod, Paris,
1967.
5.24 Lord W., Nath S., Shin Y.K. and You Z.: Electromagnetic methods of defect detection,
IEEE Trans. Magn., 1990, MAG·26, (5), pp. 2070-2075.
5.25 Marrocco A. and Pironneau 0.: Optimum design with Lagrangian finite elements: design
of an electromagnet, Comput. Methods Appl. Mech. Eng., 1978, IS, 277-308.
5.26 Michalski A., Sikora J. and Wincenciak S.: Optimal shape design of electromagnetic flow
gauge, IEEE Trans. Magn., 1988, MAG-24, (1), pp.565-568.
5.27 Mittra M.: Computer Techniques for Electromagnetics, Pergamon Press, Oxford, 1973.
5.28 Miyata K., Tawara Y., Matsui Y. and Ohashi K.: Optimum design for magnetic resonance
imaging circuit using permanent magnets, COMPEL, 1990, 9, Suppl. A, pp. 115-118.
5.29 Mustarelli P., Rudnicki M., Savini A., Savoldi F. and Villa M.: Synthesis of magnetic
gradients for NMR tomography, Magn. Reson. Imaging, 1990, 8, pp. 101-105.
5.30 Nakata T. and Takahashi N.: New design method of permanent magnets by using the finite
element method, IEEE Trans. Magn., 1983, MAG·19, (6), pp.2494-2497.
5.31 Nakata T., Takahashi N., Fujiwara K., Kawashima T. and Morii A.: Optimal design of
injection mold for plastic bonded magnet, IEEE Trans. Magn., 1991, MAG-27, (6),
pp. 4992-4994.
5.32 Osama M. Mohammed: Optimal design method of magnetic circuit by finite elements and
dynamic search procedure, COMPEL, 1990, 9, Suppl. A, pp. 107-110.
5.33 Palka R.: Application of finite element technique to continuation problems of stationary
fields, IEEE Trans. Magn., 1983, MAG·i9, (6), pp.2356-2359.
5.34 Palka R.: Extrapolation of static fields measurements by means of the finite element
method, Arch. Elektrotech., 1984, 67, pp.247-251.
5.35 Palka R.: Synthesis of electrical field by optimization of the shape of region boundaries,
Proc. lEE A, 1985, 132, pp.28-32.
5.36 Palka R.: Synthesis of magnetic field by optimization of the shape of current areas, Arch.
Elektrotech., 1989, 72, pp.293-3oo.
5.37 Palka R.: Synthesis of magnetic field due to the direct current, ETZ Archiv, 1985, 7,
pp. 299-302.
5.38 Palka R. and Sikora R.: Reverse proNem of penetration of magnetic field into conducting
region, IEEE Trans. on Magn., 1978, MAG·14, pp. 566-568.
5.39 Pawluk K.: Synthesis of electromagnetic fields, Analiza i synteza pol elektromagnetycznych,
(ed. J. Turowski), Ossolineum, Wroclaw, 1990 pp. 201-276 (in Polish).
5.40 Pawluk K.: 3-D magnetic field of coils with an open metallic core in boundary-integral
approach, Boundary Element Technology VllI, (eds: H. Sina and C. Brebbia), Computa-
tional Mechanics Publications, Southampton, Boston, 1993, pp. 147-156.
5.41 Pawluk K. and Rudnicki M.: The state of art in the synthesis of electromagnetic fields,
Electromagnetic Fields in Electrical Engineering, (eds: J. Turowski and A. Savini),
Plenum Press, New York, 1988, pp.287-292.
References 363
5.42 Phillips D.L.: A technique for the numerical solution of certain integral equation of the
first kind, J. Assoc. Compo Mech., 1962, 9-1, pp.84-97.
5.43 Preis K., Biro 0., Friedrich M., Gottvald A. and Magele c.: Comparison of different
optimization strategies in the design of electromagnetic devices, IEEE Trans. Magn.,
1991, MAG·27, (5), pp.4154-4157.
5.44 Preis K., Magele C. and Biro 0.: FEM and evolution strategies in the optimal design of
electromagnetic devices, IEEE Trans. Magn., 1990, MAG·26, (5), pp.2181-2183.
5.45 Preis K. and Ziegler A.: Optimal design of electromagnetic devices with evolution strategies,
COMPEL, 1990,9, Suppl. A, pp. 119-122.
5.46 Ratnajeevan S. and Hoole H.: Optimal design inverse problems and parallel computers,
IEEE Trans. Magn., 1991, MAG·27, (5), pp.4146-4149.
5.47 Ratnajeevan S., Hoole H. and Sirikumaran S.: Reflections off aircraft and the shape opti-
mization of a ridged waveguide, IEEE Trans. Magn., 1991, MAG·27, (5), pp. 4150-4153.
5.48 Ratnajeevan S., Hoole H. and Subramaniam S.: Higher finite element derivatives for the
quick synthesis of electromagnetic devices, IEEE Trans. Magn., 1992, MAG·28, (2),
pp. 1565-1568.
5.49 Ratnajeevan S., Hoole H. and Subramaniam S.: Inverse problems with boundary elements:
synthetising a capacitor, IEEE Trans. Magn., 1992, MAG·28, (2), pp. 1529-1532.
5.50 Rudnicki M.: Synthesis of electromagnetic field using generalized discrepancy, Pro Inst.
Elektrotech., 1985, 137, pp. 19-33.
5.51 Rudnicki M., Savini A., Di Barba P. and Pawluk K.: Optimal synthesis of electric devices
using the boundary element method, Computational Modelling of Free and Moving
Boundary Problems - 2, (eds: L.C. Wrobel and C.A. Brebbia), Walter de Gruyter, Berlin,
New York, 1988, pp. 269-279.
5.52 Russenschuck S.: Application of Lagrange multiplier estimation to the optimization of
permanent magnet synchronous machines, IEEE Trans. Magn., 1992, MAG·28, (2),
pp. 1525-1528.
5.53 Saldanha R., Coulomb J.L. and Sabonadiere J.C.: An ellipsoidal algorithm for the
optimum design of magnetostatic problems, IEEE Trans. Magn., 1992, MAG·28, (2),
pp. 1573-1576.
5.54 Saldanha R., Coulomb J.L. and Sabonadiere J.C.: A dual method for constrained
optimization design in magnetostatic problems, IEEE Trans. Magn., 1991, MAG·27, (5),
pp.4136-4141.
5.55 Salon S.J. and Istfan B.: Inverse non-linear finite element problems, IEEE Trans. Magn.,
1986, MAG·22, (5), pp. 817-818.
5.56 Savini A., Di Barba P. and Rudnicki M.: On the optimal design of air-cored solenoid
inductors of rectangular cross-section, COMPEL, 1992, 11, I, pp.205-208.
5.57 Se-yun Kim, Hyun-chul Choi, Jung-woong Ra and Soo-young Lee: Electromagnetic
imaging of 2-D inhomogeneous dielectric objects by an improved spectral inverse tech-
nique, IEEE Trans. Magn., 1990, MAG·26, (2), pp.634-637.
5.58 Siebold H., Hiibner H., SOldner L. and Reichert T.: Performance and results of a computer
program for optimizing magnets with iron, IEEE Trans. Magn., 1988, MAG·24, (I),
pp.419-422.
5.59 Sikora J.: Minimax approach to the optimal shape design of electromagnetic devices,
COMPEL, 1990,9, Suppl. A, pp.67-81.
5.60 Sikora J.: Sensitivity approach to the optimal shape design of a magnetic pole contour,
Arch. Elektrotech., 1989, 72, pp.27-32.
5.61 Sikora J. Skoczylas J., Sroka J. and Wincenciak S.: The use of the singular value
decomposition method in the synthesis of the Neumann's boundary problem, COMPEL,
1985, 4, pp. 19-27.
5.62 Sikora R. and Palka R.: Reverse problems for diffusion equation, Arch. Elektrotech., 1980,
62, pp. 177-180.
5.63 Sikora R. and Palka R.: Synthesis of one- and two-dimensional field, Arch. Elektrotech.,
1981, 64, pp. 105-108.
5.64 Sikora R., Purczynski J. and Adamiak K.: The magnetic field synthesis on a cylinder
364 References
solenoid axis by means of Tichonov's regularization method, Arch. Elektrotech., 1978,

60, pp. 83-86.
5.65 Simkin J. and Trowbridge C.W.: Optimizing electromagnetic devices combining direct
search methods with simulated annealing, IEEE Trans. Magn., 1992, MAG·28, (2),
pp. 1545-1548.
5.66 Simkin J. and Trowbridge C.W.: Optimization problems in electromagnetics, IEEE Trans.
Magn., 1991, MAG·27, (5), pp.4016-4019.
5.67 Takahashi N., Nakata T. and Uchiyama N.: Optimal design method of 3-D nonlinear
magnetic circuit by using magnetization integral equation method, IEEE Trans. Magn.,
1989, MAG·2S, (5), pp.4144-4146.
5.68 Tichonov A.N. and Arsenin W.J.: Solution Methods of Non-Correct Problems, Nauka,
Moskva, 1974 (in Russian).
5.69 Twomey S.: On the numerical solution of Fredholm integral equations of the first kind by
the inversion of the linear system produced by quadrature, J. Assoc. Compo Mech., 1963,
10-1, pp.97-101.
5.70 Twomey S.: The application of numerical filtering to the solution of integral equations
encountered in indirect sensing measurements, J. Franklin Inst., 1965,2792, pp. 95-101.
5.71 Yin Liu, Lian Gong, Ping Liu and Keqian Zhang: An algorithm and simulation of the
inverse problem of low-frequency current field, IEEE Trans. Magn., 1990, MAG·26, (2),
pp. 630-633.
Chapter 6
6.1 Brebbia C.A.: Progress in Boundary Element Method, Vol. 1, Pentech Press, London,
1981.
6.2 Ceolini F. and Lupi S.: The mutually-coupled circuits method for calculations regarding
inductors for induction heating with special configurations, IEEE lAS Conf. Proc.,
Toronto, Canada, 1-5 October, 1978, pp. 1151-1157.
6.3 Dobrogowski J. et al.: Stamping of Metals using Impulse Magnetic Field, PWN Warsaw-
Poznan, 1979 (in Polish).
6.4 Essays on the Formal Aspects of Electromagnetic Theory, (ed: A. Lakhtakia), World
Scientific. Singapore, London, 1993.
6.5 Gieras J.: Linear Induction Drives, Oxford University Press, 1993.
6.6 Gierczak E., Fleszar J. and Turowski J.: Current asymmetry in the three-phase linear motor,
Proc. ICEM'82, Budapest. 1982, Paper LM1I6, pp.961-%3.
6.7 Hammond P.: Energy Methods in Electromagnetism, Clarendon Press, Oxford, 1981.
6.8 Jackson J.D. Classical Electrodynamics, John Wiley & Sons, 1975.
6.9 Kazmierski M., Kozlowski M., Lasocinski J., Pinkiewicz I. and Turowski J.: Hot spot
identification and overheating hazard preventing when designing a large transformer,
CIGRE 1984 Session, Paris, Rep. 12-12, pp. 1-6.
6.10 Kacki E.: Thermokinetics, WNT, Warsaw 1%7, (in Polish).
6.11 Knoepfel H.: Pulsed High Magnetic Fields, North-Holland, Amsterdam 1970.
6.12 Kom~za K., Turowski J. and Wiak S.: The experimental prediction of power losses
distribution in metal elements of electrical machines, Proc. Int. Conf. on Reliability and
Lifetime of Electrical Machines, Budapest, 1984, pp.72-76.
6.13 Krajewski W.: The method of computation of coupled electromechanical fields in metal
tubes by a magnetic impulse, Pro Inst. Elektrotech., 1984, 132, pp.71-84.
6.14 Lavers J.: Numerical Solution Methods for Electroheat Problem, COMPUMAG, Santa
Margherita Ligure, 1983.
6.15 Lawrynowicz J.: Variation Calculus with Introduction to Mathematical Programming,
WNT, Warsaw 1977, (in Polish).
6.16 Mankin E.A. and Morozov D.N.: Stray losses in large power transformer cores, ClORE
1964 Session, Paris, Rep. 125.
6.17 Mendrela E.A. and Gierczak E.: Influence of primary twist on linear induction motor
performance, IEEE Trans. Magn., 1985, MAG·21, (6), pp.2664-2671.
References 365
6.18 Napieralska-Juszczak E.: The chosen problem of modeling of the electromagnetic and
thermal fields, Proc. of ISEF' 85, Warsaw, 1985.
6.19 Nenadovic V. and Riches E.E.: Maglev at Birmingham Airport, GEC Revue, No.1, 1985,
pp.3-17.
6.20 Niewierowicz N. and Turowski J.: New thermometric method of measuring power losses
in solid metal elements, Proc. lEE, 1972, 119, (5), pp.629-636.
6.21 Nowacki W.: Electromagnetic Effects in Deformed Solid Bodies, PWN, Warsaw, 1983 (in
Polish).
6.22 Ooi B.T. and White D.C.: Traction and normal forces in the linear induction motor, IEEE
Trans., No.4, 1970.
6.23 Panofsky K. and Phillips: Classical Electricity and Magnetism, Addison-Wesley,
Cambridge, Mass, 1942.
6.24 Parcus H.: Magnetothermoelasticity, CISM Unide, Springer-Verlag, Wien, 1972.
6.25 Rawa H.: Limitations following from application of Maxwell's equations in their classical
form to the description of electromagnetic fields in non-uniform media, Arch. Elektro-
tech., 1985, 34, (314), pp.765-775 (in Polish).
6.26 Reichert K.: Numerical procedure at the calculation of devices for induction heating,
Elektrowaerme Int., 1%8, 26, pp. 113-123 (in German).
6.27 Reichert K.: Sci. Electr., 1970, 16, pp. 126-146.
6.28 Sajdak C.: Electrodynamic forces in liquid metal with inductive mixing in the processes of
semi-continuous or continuous casting of cylindrical ingots, Rozpr. Elektrotech., 1985,
31, (2), pp. 423-436 (in Polish).
6.29 Sajdak C.: Influence of non-uniform distribution of speed of liquid metal on electrodynamic
forces in mixers and induction pumps, Rozpr. Elektrotech., 1986, 32, (2), pp.455-470
(in Polish).
6.30 Turowski J.: Technical Electrodynamics, WNT, Warsaw, 1993, (in Polish).
6.31 Turowski J.: Generalized equations for electromechanical and electrodynamic energy
conversion in transformers, Rozpr. Elektrotech., 1980, 26, (4), pp.911-927.
6.32 Turowski J.: Electromagnetic Computation ofElements ofElectrical Machines and Devices,
WNT, Warsaw, 1982, (in Polish).
6.33 Turowski J., Zakrzewski K., Sikora R. and Pawluk K.: Analysis and Synthesis of
Electromagnetic Fields, Polish Academy of Science. Ossolineum, Wroclaw -Warszawa
-Krakow -Gdansk -Lodz, 1990, (in Polish).
6.34 White D.C. and Woodson H.H.: Electromechanical Energy Conversion, J. Wiley, NY, 1959.
6.35 Voldek A.I.: Induction Magnetohydrodynamic Machines with Liquid Metallic Secondary,
Energia, Leningrad, 1970, (in Russian).
6.36 Zakrzewski K. and Pietras G.: Method of calculating the electromagnetic field and power
losses in ferromagnetic materials, taking into account magnetic hysteresis, Proc. lEE,
1971, (11), pp. 1679-1685.
Chapter 7
7.1 ANSYS News, Fourth Issue, Swanson Analysis Systems, Inc., Houston, 1991.
7.2 BEASY Engineering Analysis, BEASY Computational Mechanics, Ashurst Lodge,
Southampton, 1993.
7.3 Biddlecombe C.S., Sykulski J.K. and Taylor S.C.: Design Environment Modules for non-
specialist users of EM software, Record of the 9th Compumag Conference on the
Computation of Electromagnetic Fields, Miami, Florida October 31-November 4, 1993,
pp. 222-223.
7.5 Biro O. and Richter R.: CAD in Electromagnetism, Adv. Electron. Electron Phys., 1991,
82, Academic Press.
7.6 Brebbia C.A. and Dominguez J.: Boundary Elements, Computational Mechanics Publica-
tions, Southampton, 1992.
366 References
7.7 Brebbia C.A. and Aliabadi M.H. (eds): Adaptive Finite and Boundary Element Methods,
Computational Mechanics Publication, Southampton, 1992.
7.8 Bryant C.F., Emson C.R., Diserens N.J., Molinari G. and Trowbridge C.W.: Parallel
Processing and the Integration ofAnalysis and Design in Electromagnetics Computation,
COMPEL, Volume 9, Supplement A, James & James Science Publishers, London, 1990,
pp. 171-176.
7.9 Cavendish J.C. et al.: An approach to automatic 3d fe mesh generation, IJNME, 1985,21,
p.329.
7.10 Cheng Y.B. and Sykulski J.K.: A design shell for force calculations of solenoid actuators,
in Software Applications in Electrical Engineering, Computational Mechanics Publica-
tions, Southampton, 1993, pp. 39-46.
7.11 Cheng Y.B., Sykulski J.K. and Stoll R.L.: Force optimization in dc actuators, Proc. Int.
Symp. Electromagnetic Fields in Electrical Engineering, Warsaw, September 16-18, 1993,
pp.231-234.
7.12 COMPUMAG'87-Graz, Proceedings of the Conference on the Computation of Electro-
magnetic Fields, IEEE Trans. Magn., 1988, 24, (I).
7.13 COMPUMAG'89-Tokyo, Proceedings of the Conference on the Computation of Electro-
magnetic Fields, IEEE Trans. Magn., 1990, 26, (2).
7.14 COMPUMAG'91-Sorrento, Proceedings of the Conference on the Computation of Electro-
magnetic Fields, IEEE Trans. Magn., 1992, 28, (2).
7.15 COMPUMAG'93-Miami, Record of the 9th Compumag Conference on the Computation
of Electromagnetic Fields, Miami, Florida October 31-November 4, 1993.
7.16 Delaunay B.: Sur la Sphere Vide, lzvestiya Akademii Nauk, USSR, Math. and Nat. Sci.
Div., No.6, 1934, p. 793.
7.17 Drago, Molfino P., Nervi M. and Repetto M.: A 'local field error problem' approach for
error estimation infinite element analysis, IEEE Trans. Magn., 1992, 28, (2), pp. 1743-
1746.
7.18 Duff 1.S., Erisman A.M. and Reid J.K.: Direct Methods for Sparse Matrices, Clarendon
7.19 Dyck D.N., Lowther D.A. and McFee S.: Determining an approximate finite element mesh
density using neural network techniques, IEEE Trans. Magn., 1992, 28, (2), pp. 1767-
1770.
7.20 Fernandes P., Girdinio P., Molfino P. and Repetto M.: Local error estimates for adaptive
mesh refinement, IEEE Trans. Magn., 1988, 24, (I), pp.299-302.
7.21 Fernandes P., Girdinio P., Molfino P., Molinari G. and Repetto M.: A comparison of
adaptive strategies for mesh refinement based on 'a posteriori' local error estimation
procedures, IEEE Trans. Magn., 1990, 26, (2), pp. 795-798.
7.22 Fernandes P., Girdinio P., Molinari G. and Repetto M.: Local error estimation proce-
dures as refinement indicators in adaptive meshing, IEEE Trans. Magn., 1991, 27, (5),
pp.4189-4192.
7.23 Fernandes P., Girdinio P., Repetto M. and Secondo G.: Refinement strategies in adaptive
meshing, IEEE Trans. Magn., 1992, 28, (2), pp. 1739-1742.
7.24 Fletcher R.: Practical Methods of Optimization, 2nd edn, John Wiley & Sons, 1987.
7.25 Golias N.A. and Tsiboukis T.D.: Adaptive refinement in 2-D finite element applications,
Int. J. Numer. Model. Electron. Netw. Devices Fields, 1991, 4, pp. 81-95.
7.26 Golias N.A. and Tsiboukis T.D.: A-Posteriori Adaptive Mesh Refinement In the Finite
Element Eddy Current Computation, COMPEL, Volume 11, No.1, James & James
Science Publishers, London, 1992, pp. 249-252.
7.27 Golias N.A. and Tsiboukis T.D.: Three-dimensional automatic adaptive mesh generation,
IEEE Trans. Magn., 1992, 28, (2), pp. 1700-1703.
7.28 Golias N.A. and Tsiboukis T.D.: Adaptive refinement strategies in three dimensions, IEEE
Trans. Magn., 1993, 29, (2), pp. 1886-1889.
7.29 Gramz M. and Zi61kowski M.: SONMAP V.2.0, User Manual, Szczecin, 1988 (in Polish).
7.30 Hammond P. and Sykulski J.K.: Engineering Electromagnetism: Physical Processes and
Computation, Oxford University Press, Oxford, 1994.
References 367
7.31 Hammond P. and Sykulski J.K.: Tubes and slices: a new way of teaching the principles
of electric and magnetic fields, IEEE Trans. Educ., 1992, 35, (4), pp.300-306.
7.32 Hoole S.R.H.: Nodal perturbations in adaptive expert finite element mesh generation, IEEE
Trans. Magn., 1987, 23, pp.2635-2637.
7.33 Ida N.: Finite Element Computation on SIMD and MIMD Parallel Computers, COMPEL,
Volume 9, Supplement A, James & James Science Publishers, London, 1990, pp. 177-
180.
7.34 Integrated Engineering Software, Winnipeg, Canada, 1993.
7.35 International Compumag Society, Newsletter, October 1993, I, (I).
7.36 ISEF'93, Proc. Int. Symp. Electromagnetic Fields in Electrical Engineering, Warsaw,
September 16-18, 1993.
7.37 Jin J.M.: The Finite Element Method in Electromagnetics, John Wiley & Sons, New York,
1993.
7.38 Lindholm D.A.: Automatic mesh generation on surfaces of polyhedra, IEEE Trans. Magn.,
1983, 19, (6), pp. 2539-2542.
7.39 Long W., Piriou F. and Razek A.: An Adapted Cholesky Decomposition Method/or the
Solution 0/ Coupled Magnetic Electric Equations, COMPEL, Volume 8, No.4, James
& James Science Publishers, London, 1989, pp.203-208.
7.40 Lowther D. and Silvester P.: Computer-Aided Design in Magnetics, Springer-Verlag, New
York, 1986.
7.41 MagNet 5.0, User Manual, Infolytica Limited, Swindon, 1993.
7.42 Magnetics Update, Newsletter, 1993, 5, (3), Infolytica Corporation, Montreal, Canada.
7.43 Magnetics Update (Europe), 1993, 1, (I), Newsletter, Infolytica Ltd, London, UK.
7.44 MATLAB and Optimization Toolbox, User Manual, The MathWorks, Inc., California,
1994.
7.45 Meijerink J.A. and der Vorst V.: An iterative solution method for systems of which the
coefficient matrix is a symmetric in matrix, Math Comp, 1977, 31, p. 148.
7.46 Marchand C. and Razek A.: Optimal torque operation of digitally controlled permanent
magnet synchronous motor drives, lEE Proc. A, 1993, 140, (3), pp. 232-240.
7.47 Nakata T. (ed.): 3-D Electromagnetic Field Analysis, COMPEL, Volume 9, Supplement A,
James & James Science Publishers, London, 1990.
7.48 OPERA, User Manual, Vector Fields Ltd, Oxford, 1993.
7.49 Pan Q., Kladas A. and Razek A.: Magneto-thermal coupled nonlinear finite element
analysis in squirrel cage induction motors, Math. Engng Ind., 1989,2, (3), pp. 203-217.
7.50 Parallel Applications Centre, University of Southampton, 1994.
7.51 PC-OPERA, User Manual, Vector Fields Ltd, Oxford, 1993.
7.52 Penman J. and Fraser J.R.: Dual and complementary energy methods in electromagnetism,
IEEE Trans. Magn., 1983, 19, (6), pp.2311-2316.
7.53 Penman J. and Grieve M.D.: An approach to self adaptive mesh generation, IEEE Trans.
Magn., 1985, 21, (6), pp.2567-2570.
7.54 Penman J. and Grieve M.D.: Self-adaptive mesh generation technique for the finite element
method, lEE Proc. A, 1987, 134, (8), pp.634-650.
7.55 Pichon L. and Razek A.: Electromagnetic field computations in a three-dimensional cavity
with a waveguide junction of a frequency standard, lEE Proc. A, 1992, 139, (4),
pp. 343-346.
7.56 Pichon L. and Razek A.: Hybrid finite-element method and boundary-element method
using time-stepping for eddy-current calculation in axisymmetric problems, lEE Proc. A,
1989, 136, (4), pp.217-222.
7.57 Pinchuk A.M. and Silvester P.: Error estimation for automatic adaptive finite element mesh
generation, IEEE Trans. Magn., 1985, 21, (6), pp. 2551-2554.
7.58 Raizer A., Meunier G. and Coulomb J.L.: An approach for automatic adaptive mesh
refinement in the finite element computation of magnetic fields, IEEE Trans. Magn.,
1989, 25, (4), pp. 2965-2%7.
7.59 Reid J.K.: On the Method o/Conjugate Gradients/or the Solution 0/ Large Sparse Systems
0/ Equations, Academic Press, New York, 1971.
368 References
7.60 Rencis 1.J. and Brebbia C.A. (eds): Boundary Elements XV, Computational Mechanics
Publications, Southampton, 1993.
7.61 Rikabi J., Bryant C.F. and Freeman E.M.: An error-based approach to complementary
formulations of static field solutions, Int. J. Numer. Methods Eng., 1988,26, pp. 1963-
1987.
7.62 Russenschuck: Mathematical optimization techniques for the design of permanent magnet
synchronous machines based on numerical field calculation, IEEE Trans. Magn., 1990,
26, (2), pp. 638-641.
7.63 Russenschuck: Application of Lagrange multiplier estimation to the design optimization of
permanent magnet synchronous machines, IEEE Trans. Magn., 1992,28, (2), pp. 1525-
1528.
7.64 Silvester P. and Ferrari R.: Finite Elements for Electrical Engineers, Cambridge University
Press, Cambridge, 1990.
7.65 Silvester P. (ed.): Advances in Electrical Engineering Software, Computational Mechanics
Publications, Southampton, 1990.
7.66 Silvester P. (ed.): Software Applications in Electrical Engineering, Computational
Mechanics Publications, Southampton, 1993.
7.67 Silvester P.: Data Structures for Engineering Software, Computational Mechanics Publica-
tions, Southampton, 1993.
7.68 Simkin J. and Trowbridge C.W.: Optimization problems in electromagnetics, IEEE Trans.
Magn., 1991, 27, (5), pp.4016-4019.
7.69 Simkin J. and Trowbridge C.W.: Optimizing electromagnetic devices combining direct
search methods with simulated annealing, IEEE Trans. Magn., 1992, 28, (2), pp. 1545-
1548.
7.70 Sloan S.W.: A fast algorithm for constructing Delaunay triangulations in the plane,
Advanced Engineering Software, Ch. 9, 1987, pp. 34-55.
7.71 Sykulski J.K. and Turowski J. (eds): ISEF'91 Proceedings, COMPEL, Volume 11, No. I,
James & James Science Publishers, London, March 1992.
7.72 Sykulski J.K.: Computer package for calculating electric and magnetic fields exploiting dual
energy bounds, lEE Proc. A, 1988, 135, (3), pp. 145-150.
7.73 Sykulski J.K. and Stoll R.L.: Magnetic field modelling and calculation of reflected
impedance of inductive sensors, IEEE Trans. Magn., 1992, 28, (2), pp. 1426-1429.
7.74 Sykulski J.K. and Stoll R.L.: Finite Element Modelling of Inductive Sensors, COMPEL,
Volume 11, No. I, James & James Science Publishers, London, 1992, pp.69-72.
7.75 Sykulski J.K., Sykulska E. and Hughes S.T.: Application of Finite Element Modelling in
LVDT Design, COMPEL, Volume 11, No. I, James & James Science Publishers,
London, 1992, pp.73-76.
7.76 Sykulski J.K.: Tubes and Slices: a software package for dual field calculations, in Soft-
ware Applications in Electrical Engineering, Computational Mechanics Publications,
Southampton, 1993, pp.247-254.
7.77 Trowbridge C.W.: The Work in Progress: Main Development Trends in Electromagnetic
Analysis, Technical Course: Field Computation for Low Frequency Devices, CEMM
(lEE) course, Southampton, June 1993.
7.78 Tsiboukis T.D.: Automatic finite element analysis, Private communication, 1993.
7.79 Vector, 1993, 9, (2), Electromagnetics Newsletter, Vector Fields Ltd, Oxford, UK.
7.80 Ziolkowski M.: 3D scalar or vector magnetic field - how one can see it, ISEF'93, Proc.
Int. Symp. Electromagnetic Fields in Electrical Engineering, Warsaw, September 16-18,
1993, pp. 273-276.
Chapter 8
8.1 Bland T.G. and Freeman E.M.: Scale-model linear induction motors with solid iron
secondaries, Proc. lEE, 1978, 125, (11), pp. 1223-1226.
8.2 Boriu N.W.: Modelling of leakage power loss of high power transformers, Elektrichestvo,
1960, (9), pp. 38-42 (in Russian).
References 369
8.3 Dobkowski J.: External leakage fluxes of DC machines, Technical University of Lodz,
Ph.D. thesis, 1977 (in Polish).
8.4 Filtshakov P.F. and Pantshishin B.N.: EGDA integrators. Potential field modelling on
Semi-Conducting Paper, Kiev, 1961 (in Russian).
8.5 Freeman E.M., Lowther D.A. and Laithwaite E.R.: Scale-model linear induction motors,
Proc. lEE, 1975, 122, (7), pp.721-726.
8.6 Gilbert A.J.: A method of measuring loss distribution in electrical machines, Proc. lEE,
1961, lORA, pp. 239-244.
8.7 Govorkov W.A.: Electric and Magnetic Fields, WNT, Warsaw, 1984 (in Polish).
8.8 Hagel R. and Zakrzewski J.: Dynamic Measurements, WNT, Warsaw, 1962 (in Polish).
8.9 Hippner M.: Graphical Presentation of Magnetic Field Calculation Results, Z.N. ATR,
Bydgoszcz (in Polish).
8.10 Ivanow-Smolensky A.W.: Electromagnetic Fields and Phenomena in Elecrical Machines,
and their Modelling, Energia, Moscow, 1969 (in Russian).
8.11 Jablonski M.: Transformer Tests in Industry and Exploitation, WNT, Warsaw, 1969 (in
Polish).
8.12 Komeza K., Turowski J. and Wiak S.: The experimental prediction of power loss dis-
tribution in metal elements of electrical machines, Proc. of ICEM' 84, Budapest, 1984,
pp.72-76.
8.13 Koziowska A.: Measurements and analysis of flux density distribution in 3-phase trans-
former core made of grain oriented steel, Rozpr. Elektrotech., (2), 1967, p.347 (in
Polish).
8.14 Laczkowski R.: Vibroacoustic of Machines and Devices, WNT, Warsaw, 1983 (in
Polish).
8.15 Nalecz M. and Jaworski J.: Magnetic Measurements, PWN, Warsaw-Lodz, 1965 (in
Polish).
8.16 Neyman L.R.: Skin-Effect in Ferromagnetic Materials, GEl, Leningrad-Moscow, 1949 (in
Russian).
8.17 Niewierowicz N. and Turowski J.: New thermometric method of measuring power losses
in solid metal elements, Proc. lEE, 1972, 119, (5), pp.629-636.
8.18 Nowaczynski J.: Local overheating of construction parts of transformers during over-
loading and overexcitation, Technical University of Lodz, PhD thesis, 1977 (in Polish).
8.19 Pietrzak S. and Zakrzewski K.: Measurement system of static characteristics of double-
sided linear motor, Z.N. Elektryka, 1983, (74), pp. 337-345 (in Polish).
8.20 Styburski W.: Tensometric transducers, construction, design, exploitation, WNT, Warsaw,
1976 (in Polish).
8.21 Szumilewicz B., Slomski B. and Styburski W.: Electronic Measurements in Engineering,
WNT, Warsaw, 1982 (in Polish).
8.22 Turowski J.: Technical Electrodynamics, WNT, Warsaw, 1968 (in Polish).
8.23 Turowski J., Kazmierski M. and Ketner A.: Thermometric method for power loss measure-
ments in construction parts of transformers, Prz. Elektrotech., (10), 1964, pp.439-448
(in Polish).
8.24 Venikov W.A. and Ivanov-Smolensky A.W.: Physical Modelling of Electrical Systems,
GEl, Moscow, 1956 (in Russian).
8.25 Zakrzewski K.: Practical method of calculating active and reactive power in solid iron,
Rozpr. Elektrotech., 1975, 21, (1), pp. 215-233 (in Polish).
8.26 Zakrzewski K.: Electromagnetic field in ferromagnetic materials, Z.N. Elekryka, (38), 1972
(in Polish).
8.27 Zakrzewski K. Physical modelling of field and loading loss in transformers, Rozpr.
Elektrotech., 1979, 25, (2), pp.401-418. (in Polish).
8.28 Zakrzewski K. Scale modelling of linear induction motors taking into account magnetic
nonlinearity of iron secondaries, Rozpr. Elektrotech., 1983, 29, (2), pp.469-489.
8.29 Zakrzewski K.: Physical modelling of leakage field and stray losses in steel constructional
parts of electrotechnical devices, Arch. Elektrotech., 1986, (69), pp. 129-135.
8.30 Zakrzewski K. and Pietras F.: Method of calculating the electromagetic field and power
370 References
losses in ferromagnetic materials taking into account magnetic hysteresis, Proc. lEE,
1971, 118, (11), pp. 1679-1785.
8.31 Zakrzewski K. and Sykulski J.: The effect of eddy currents on physical modelling of
magnetic field strength at increased frequencies, Rozpr. Elektrotech. 1981, 72, (1),
pp.53-63.
8.32 Zakrzewski K.: Mathematical modelling of electromagnetic field in solid iron, Rozpr.
Elektrotech., 1970, 16, (I), pp. 27-43 (in Polish).
Index
adaptive meshing 287, 305-13 diamagnetism 13

Alternating-Direction Implicit differential forms 21-4
method (ADI) 72 diffusion equation 9, 38, 40, 44,
Ampere's law 34, 19, 76, 245 49
analogue modelling 348-52 Dirichlet condition 12, 56, 79, 86
discretization error 42
boundary collocation method 81 div operator 2
boundary conditions DuFort-Frankel algorithm 41, 50,
electric 10 56, 70-2
external 11
magnetic 9-12 eddy current 18-20, 49-52, 320
boundary-element method (BEM) eigenfunction 184
109-12 eigenvalue 184
bounds, upper and lower 24-6 electric vector potential 76
electromagnetic screening 158
CAD 285-329 energy density 15, 241
Cholesky decomposition 303-5 Epstein's apparatus 275, 341
coenergy 15-17, 236 error analysis 305-13, 353
coercive force 14 Euler-Lagrange's equation 241
complimentary formulation 308 experimental methods 330-53
COMPUMAG conference 285,297 exterior differential operator
conductivity 2 21-3
conjugate gradient method 303-5,
314 Faraday's law 4
constitutive equations 2, 262, 309 ferromagnetism 13, 52
continuity equation 1 finite-difference method 35-77
convergence 42-8 finite-element method (FEM)
Coulomb gauge 6 78-144
coupled fields 234-84, 345-8 force 15-18, 61
Crank-Nicolson algorithm 40, 102, Fourier series 36
107 Fredholm's equation 183-6
curl operator 2
Galerkin method 80-2
Delaunay triangulations 289-95 gauge transformation 7
OEM 288 Gauss's theorem 3
372 Index
Gauss-Seidel iterative algorithm magnetothermoelasticity 265

61, 135 Maxwell's equations 1, 244, 330
gradient operator 2 Maxwell's stress 18, 252, 315
Green's functions 186-198 MHD 254
Green's theorem 4, 105, 109
nabla operator 3
Hamilton's principle 237 Neumann condition 12, 56, 58,
Helmholtz equation 8, 9, 18, 63, 79, 86, 114
68, 86, 138, 188 Newton-Raphson method 94, 107
hierarchical elements 123-5
Hodge operator 21-4 open-boundary problems 113-23
Hopscotch algorithm 41 OPERA program 289, 316, 326-7
hysteresis 14, 53 optimization 313-5
ICCG method 303-5 parabolic equation 9, 37-42

inductance 60 parallel computing 321-3
induction heating 276 paramagnetism 13
interactive design 160, 164 permeability 2
interface conditions 9 differential 14, 39
inverse problems 180, 196-201 incremental 14
ISEF conference 285 reversible 14
isoparametric elements 90- 3 permeance 29
permittivity 2
Jacobian matrix 94, 97 physical modelling 330-338
point collocation method 80
kernel 183 Poisson's equation 9, 57, 73-7,
148, 172, 199
Lagrangian 237 post-processor 147, 286, 316-21
Laplace's equation 8, 73-7, 138, Poynting vector 244, 337, 340
196 pre-processor 147, 286, 289-96
leakage field 137, 150-60, 344
least action principle 238 quadrature 204
least-squares method 80, 214, 314
levitation 281 regularization 219-24
linear motor 283, 335 reluctance network 145-78
Lorentz gauge 6, 8, 22 remanent flux density 14
Lorentz force 247, 254 Reynolds number, magnetic
LVDT 296, 317 257-60
RNM program 147, 152, 161
MagNet program 324-6
magnetic screening 158 scalar potential 4-8
magnetization vector 12 scaling criteria 330-8
magnetoelasticity 261 skin depth 19
magnetomotive force (mmf) 15, skin effect 18-20
67, 173 solenoid actuator 316
magnetostriction 261 SOR 61-2
Index 373
sparse matrix 88-90, 298-303 transformer 58-60, 137-44,

stability 42 150-71, 343
stiffness matrix 84 transient solutions 55, 70, 101-9,
Stoke's theorem 3, 63, 68 174-6
surface impedance 20 triangular elements 89
surface polarity 11 triangular decomposition 132
susceptibility, magnetic 12 truncation error 37, 40
synthesis of field 179-233 tubes and slices 24-33
TAS program 29-33 vector potential 5-8

Taylor series 64-5 virtual work principle 236, 247,
TEAM workshop 171, 177, 297 315
tetrahedral element 126, 307 Volterra's equation 182-6
thermal fields 261
thermokinetics 267 wave equation 9
thermometric method 347 weighted residuals 78-82

Jan K. Sykulski, Richard L. Stoll (Auth.), Jan K. Sykulski (Eds.) - Computational Magnetics-Springer Netherlands (1995)

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Jan K. Sykulski, Richard L. Stoll (Auth.), Jan K. Sykulski (Eds.) - Computational Magnetics-Springer Netherlands (1995)

Uploaded by

Copyright:

Available Formats

Computational Magnetics

© 1995 Springer Science+Business Media Dordrecht

ISBN 978-94-010-4557-5 ISBN 978-94-011-1278-9 (eBook)

(^) Printed on permanent acid-free text paper, manufactured in

accordance with ANSI/NISO Z39.48-1992 and ANSI/NISO Z39.48-1984

1. A Brief Review of Magnetic Fields

2.6.6 Two-dimensional transient or nonlinear

3. Finite-Element and Boundary-Element Methods 78

4. Reluctance Networks 145

5. Field Synthesis 179

5.3.1 Inverse problem to Laplace's equation 196

6, Coupled Fields 234

7. Computer Aided Design in Magnetics 285

8. Experimental Methods 330

Computational Magnetics comprises a spectrum of subjects covering field and

magnetics as covered by a typical second year undergraduate course. A

presents a rather unique implementation of an old approach based on equi-

• Krystyn Pawluk - Professor of Electrical Engineering, Institute of Electro-

• Richard L. Stoll - Reader in Electrical Machines, Department of Electrical

The collaboration between the Department of Electrical Engineering at

1.1 MAXWELL'S EQUATIONS

divJ = _ iJp (1.5)

The above equations are supplemented by the constitutive equations

curl F = lim ~ " F . dl (1.17)

and the three operations are reduced to

and Stoke's or circulation theorem

It is also important to notice that the differential formulation of Faraday's

HI (uv v - vv U) dv = ~ (UVV - VVU) . ds

which are particularly helpful in the formulation of the finite-element and

1.2 SCALAR AND VECTOR POTENTIALS

F = - grad t/> + curl A (1.34)

V 2t/> = _.': (1.42)

which is a second-order differential equation, known as the Poisson 's

In conductors it is likely that 11 = 110' so that

For convenience we are allowed to put

~' = ~ _ a", (1.61)

V2~ _ k2~ =_g (1.73)

1.3 CLASSIFICATION OF FIELDS

1.4 BOUNDARY CONDITIONS

Boundary conditions at an interface

where the contribution of the ribbon edge of the coin-shaped volume is

from which qm follows in terms of H n. Region 1 is assumed here to be the

qm = (1l,2 - 1l,1) ~ H nl (1.95)

External boundary conditions

problems numerically may require special techniques. More usually a prob-

Ifs (n x H) dS = JHcurlHdV = JHJdV

In a two-dimensional problem with the current flow perpendicular to the plane

In terms of the scalar magnetic potential 4>m (eqn 1.52), HI corresponds to

1.5 MAGNETIC MATERIALS

where the relative permeability

Diamagnetic Xm is a very small negative number (J.1, < 1),

Diamagnetism and paramagnetism are very small effects, with Xm typically of

Fig. 1.1 Hysteresis loops for a ferromagnetic material

1.6 ENERGY STORAGE AND FORCES IN MAGNETIC

The total magnetic energy stored is

If, for simplicity, we assume a uniform distribution of field in the region

Wm = J (H x length) (area x dB) = J mmj(cP) dcP (1.106)

Wm = Ji(A.) dA. (1.107)

Io BdH = I1..(;) di = I (j>(mmf) dmmf

Consider now Fig. 1.3 which shows a change of Wm at constant current

F = i [a1..J - [awmJ (1.111)

Fig. 1.3 Change of energy at constant current

F = [a(iA.)] _(awmJ = (aw,;,] (1.112)

Combining eqns (1.95) and (1.113) we also obtain